# Your Moat Is Your Data Model

Mike Phipps

AI Engineer World's Fair 2026

[Talk page](https://ai.engineer/talks/jt1Pbr_n6oU-your-moat-is-your-data-model) · [Video](https://www.youtube.com/watch?v=jt1Pbr_n6oU)

## Your Moat Is Your Data Model

The Gates Foundation’s Strategic Intelligence Platform turns operational knowledge into a graph that agents can traverse, while keeping reporting rules and organizational context explicit.

Familiarity with database relationships and AI tool calling will help; the article explains the graph structures as they appear.

### What should an AI platform team own?

[0:43](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=43s)

If tools such as Claude Code make applications faster to build, what remains defensible once those applications reach production? Shipping introduces monitoring, upkeep and downstream dependencies. The team must decide how much of that stack it wants to maintain—and whether users need another chat application when they already work in **Claude** or **ChatGPT**.

For Mike Phipps and his team, the durable asset is **knowledge of internal processes**: the tacit understanding required to interpret organizational data correctly. A better model does not automatically know what an internal field means, how a reporting convention works or which relationships matter to a business question. Encoding that knowledge gives the team something useful across successive models and interfaces.

![Slide titled “Intro: What’s defensible for your organization” lists production constraints and emphasizes internal processes and tacit knowledge, with the presenter inset at left.](https://ai.engineer/api/talk-essay-media/6b2922feeae0d0687c56a57f816a340b817fd1c73d065f92ff367f190518f1ab)

Internal processes and tacit knowledge as an organizational moat.

That work became the **Strategic Intelligence Platform**, or SIP. Phipps reports that SIP entered enterprise production in the preceding month for a population of about 4,000 people across the Gates Foundation. The platform makes operational knowledge available to AI without making a proprietary chat interface the center of the product.

### The organization the graph must represent

[2:36](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=156s)

The foundation’s work spans child mortality, nutrition, agriculture and education, with more than 25 years of accumulated operational data. Its interventions include creating market incentives, encouraging innovation and connecting public and private organizations. Extracting useful evidence from those investments requires understanding the structure around the data, not merely searching the documents it produced.

Phipps’s 2023 snapshot describes more than 2,000 grants, many above $5 million, targeting more than 100 countries. The scope also includes grantees, alumni, employees, multiple program strategies and activity across almost all U.S. states. The financial scale is substantial: the [2023 annual report](https://www.gatesfoundation.org/about/financials/annual-reports/annual-report-2023) records $7.749 billion in total charitable support, distinct from $6.827 billion in direct grantee support. The broader total should not be read as grant payments alone.

![Eight statistics show 2,396 total grants, 135 countries, 2,238 alumni, 44 program strategies, 48 U.S. states, 1,422 grantees, $7.75B total charitable support, and 2,026 employees.](https://ai.engineer/api/talk-essay-media/1556af96c24c7ded9897e0871da496238d3a8d68c3d3ce6da4518538ef3812d7)

The foundation’s scope of work, with statistics for new grants from FY 2023.

Funding flows through divisions including Global Development, Global Health, Gender Equality and the United States Program, abbreviated USP. Those divisions, their strategies and their investments are part of the structure an agent must understand when answering questions about the foundation’s work.

### From siloed records to shared meaning

[5:05](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=305s)

SIP structures operational data for **agentic retrieval**. Its knowledge graph is designed with an agent as the consumer. Traditionally separate systems of record supply structured and unstructured information to a data lakehouse, bringing internal enterprise records and programmatic outputs of investments under one roof. A curation layer processes that material before SIP exposes it through agentic chat and workflows. The result is a semantic graph that agents can reason across, rather than a collection of isolated source systems.

Consolidating storage does not settle the meaning of the records. Data owners must explain field definitions, dataset structure and valid joins. They also supply knowledge of limitations, systematic quirks, safeguards, security trimming and reporting conventions. A technically plausible answer can still fail if it uses a definition that differs from the organization’s established reports. **Correctness includes the organization’s rules for interpreting its data.**

### Prepare the data before making it easy to retrieve

[7:09](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=429s)

The curation pipeline handles several distinct responsibilities:

- **Preprocessing:** Filter and deduplicate datasets, preserve meaningful document ordering and resolve inconsistencies across documents upstream.
- **Extraction:** Extract structured fields, divide unstructured documents into semantic chunks and convert information in figures into text that retrieval can reach.
- **Graph enrichment:** Use tags to establish connections and generated metadata to populate graph properties.

These steps determine what information is available to the agent and how that information connects to the rest of the organization.

Governance becomes more consequential when AI makes information easier to find. Data that was technically accessible but difficult to discover can become immediately available through a natural-language question. Phipps therefore treats PII masking, sensitive-data classification and per-user entitlements as necessary parts of the pipeline. Existing accessibility is not sufficient evidence that a dataset is safe to expose through a much more effective retrieval interface.

### Start with annual reviews, then follow the funding

[8:23](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=503s)

The graph begins with a recognizable organizational process: annual strategy reviews. Phipps describes more than 80 strategy teams whose annual reviews determine yearly budgets. Review meetings bring unstructured documents into the model while providing structured links to other systems of record. The displayed model is conceptual: a single type on the diagram can represent many instantiated nodes, with one-to-one or one-to-many relationships in the actual graph.

The first hierarchy is funding. Phipps describes a five-level **additive directed acyclic graph**, or DAG, in which all levels matter together. Roll-ups must preserve the funding path rather than treat a single level as the complete answer. An in-path shortcut connects that path, while the relationship he calls “Funds to BAU” is where funding-team budgets are stored.

Funding teams have portfolios that support investments, and multiple teams can fund the same investment. This makes funding a many-to-many relationship, not a single owner field on an investment record. Investments are the foundation’s product; internal funds express priorities among different kinds of investments. The model can retain transaction-level detail or represent annual aggregations, depending on the question being answered.

From an investment, the graph can extend to different kinds of organizations and to observable outputs such as published reports and products. Connecting those outputs to the investment also connects them to its organizational context. Phipps identifies this as an area with substantial room still to populate, rather than a finished inventory of everything the investments have produced.

### Management is a different relationship

[11:27](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=687s)

The management hierarchy answers a different question about the same investment. Here, each organizational level has meaning independently; the essential distinction is how attribution works, rather than whether it follows the funding hierarchy’s topology. `Contains` relationships connect teams through the hierarchy. A team can manage an investment directly, while an ancestor team also receives indirect attribution for investments managed by its descendants.

SIP precomputes that indirect attribution. After creating the containment and direct-management relationships, it derives a management roll-up edge. This lets retrieval use a relationship whose organizational meaning has already been established instead of rediscovering the roll-up rule for every question.

For a small teaching example, suppose `team-parent` contains `team-child`, and `team-child` directly manages `investment-a`. With illustrative Cypher labels and relationship names, the derivation can be expressed as:

```cypher
MATCH (ancestor:Team)-[:CONTAINS*1..]->(direct:Team)
      -[:MANAGES_INV]->(investment:Investment)
WITH DISTINCT ancestor, investment
MERGE (ancestor)-[:ROLLUP_MANAGES_INV]->(investment);
```

For those records, the query would add a roll-up from `team-parent` to `investment-a`, preserving the original containment and direct-management relationships. `DISTINCT` collapses repeated ancestor–investment pairs before `MERGE` creates or matches the derived edge.

Both lenses remain available on the same investment:

| Lens | Relationship being represented | Roll-up concern |
|---|---|---|
| Funding | Teams contribute through funding paths | Preserve additive contributions |
| Management | Teams manage directly or through descendants | Attribute descendant investments to ancestors |

A question about a team’s investments can therefore require clarification: does the user mean what the team funds, what it directly manages or what falls under its management hierarchy? Keeping the relationships separate makes those interpretations representable.

### People connect the organizational views

[12:38](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=758s)

People provide a third hierarchy. Organizational charts, reporting relationships and team membership connect individuals to the operating structure. The same person can also appear as an investment owner, a meeting attendee or a director. Those roles add context that would be lost if personnel records remained isolated in an HR system.

The cross-system graph depends on finding **shared entities** in those formerly separate sources. A person, team or other organizational entity becomes a point where related records can be stitched together. Once those connections exist, an agent can traverse the organizational structure instead of treating each source as a separate body of facts.

### Documents become discoverable in context

[13:49](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=829s)

At the time of the talk, SIP had ingested one document source. Its Neo4j model connects meetings to documents, then documents to semantic sections or chunks. Full-text indexes support searching that content, while the relationships keep each result attached to its meeting and organizational context. An agent can also retrieve information through graph traversal alone.

![Graph slide highlights blue meeting and document nodes connected to dimmer organizational nodes, beside a panel titled “Documents become discoverable, in context.”](https://ai.engineer/api/talk-essay-media/ba414f3dc5a73c0a588411ad77eecb78b3205c318c7198b742f75d8f77127a97)

Documents become discoverable in the context of the organizational graph.

The complete model brings four source systems into one graph, exposed to agents as a semantic layer through the **Model Context Protocol**, or MCP. Agents can discover structure and reason across it at query time. The modeling process also benefits developers: drawing the connections quickly exposes missing datasets and gaps in the team’s own understanding of the organization.

### Deliver the graph where users already work

[15:26](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=926s)

SIP reaches users through MCP connections to the chat products they already use, including Claude and ChatGPT. The team chose not to own the chat interface and, for some general-chat use cases, not to own the agent interaction either. Its investment stays concentrated in the organizational knowledge those interfaces can access.

Phipps says the team forked an off-the-shelf Neo4j MCP server and modified it substantially. Changes included schema updates and mechanisms for passing state back to SIP, such as conversation IDs and message numbers. The [current official Neo4j MCP server](https://github.com/neo4j/mcp) is a related implementation, not a specification of SIP’s historical fork or its private tool contract.

General chat is one entry point; constrained workflows are another, still under development in the talk. Phipps describes access through Cowork or Claude Chat, with [MCP Apps](https://blog.modelcontextprotocol.io/posts/2026-01-26-mcp-apps/) bringing workflow interfaces into the chat experience. Sandbox-based agents would run the workflows against the same knowledge-graph backend. This keeps the shared data model useful while allowing a more controlled experience than an open-ended conversation.

### Evaluate against the live graph

[17:23](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1043s)

Evaluations reveal where a data model is incomplete or ambiguous. Users may ask questions that admit several interpretations, or receive answers that do not conform to established reporting standards. SIP’s team works with data owners to turn those standards into targeted evaluation questions.

The reference answer cannot simply be a frozen value because the structured data keeps changing. The evaluation procedure therefore preserves a reference query for each question:

1. **Organize questions by complexity.** Separate simpler requests from those requiring more involved reasoning.
2. **Maintain a reference graph query.** Encode the intended interpretation of each question in a query.
3. **Run it against the live graph.** Retrieve the expected result when the evaluation runs.
4. **Compare the agent’s answer.** Judge the response against that current result rather than an obsolete snapshot.

This separates changes in the underlying records from failures to interpret or retrieve them correctly.

The evaluation pipeline uses an LLM as a judge and tracks `pass@1` alongside stability: whether repeated asks of the same question produce consistent answers. Failures feed back into the data model, domain rules and schema descriptions. Evaluation is therefore a way to improve the knowledge supplied to the agent, not only a way to score its final response.

Phipps describes the system’s `pass@1` and stability as very strong, without giving numerical results in the spoken account. He says remaining misses tend to involve ambiguity: an answer can be valid under one interpretation yet differ from what the user intended. That distinction makes reporting definitions and domain rules part of the continuing engineering work, even when retrieval returns factually defensible results.

### Expand coverage without losing the shared context

[19:27](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1167s)

The next phase starts by filling gaps in records from systems already represented in SIP, then expanding the primary graph to additional enterprise datasets. Teams also want to connect their own data to the enterprise model through a federated graph experience; Phipps describes the approach as work in progress. Further agentic experiences build on that expanding foundation. The open engineering problem is how to make more of the organization’s knowledge available while preserving the shared meanings that make cross-system answers useful.

![Slide titled “What’s ahead for SIP” lists filling out existing data, expanding enterprise datasets, linking team datasets through federated graphs, and building controlled agentic workflow experiences.](https://ai.engineer/api/talk-essay-media/38b5fcbfbc87f2a62ef60fa4ef002bed60a748ef6b45dd8fcda07f0c1c5b9b40)

SIP’s roadmap: broader data coverage, federated graphs, and agentic workflows.

## Resources

- [Gates Foundation Annual Report 2023](https://www.gatesfoundation.org/about/financials/annual-reports/annual-report-2023) — The foundation's 2023 spending breakdown by division and strategy, with definitions of its financial totals.
- [Neo4j Official MCP Server](https://github.com/neo4j/mcp) — An MCP server providing graph-schema discovery and Cypher query tools for compatible AI clients.
- [Neo4j Full-text Indexes](https://neo4j.com/docs/cypher-manual/current/indexes/semantic-indexes/full-text-indexes/) — Instructions for creating and querying full-text indexes over node and relationship properties.
- [MCP Apps](https://blog.modelcontextprotocol.io/posts/2026-01-26-mcp-apps/) — The official introduction to interactive interfaces delivered by MCP tools inside chat clients.

## Slides

### Your Data Model is Your Moat

Mike Phipps introduces the Gates Foundation’s Strategic Intelligence Platform rollout and the decisions behind its enterprise-wide launch.

[0:16](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=16s) [Title slide for Mike Phipps’s presentation on the Strategic Intelligence Platform enterprise rollout.](https://assets.ai.engineer/sites/aiecode2025/sha256/ce97e296413e863acc024b126355bd5eb73678fd61cb0758b4b63d996bbeb9cc.webp)

Your Data Model is Your Moat

Strategic Intelligence Platform (SIP) Enterprise Rollout

Mike Phipps
Lead AI Engineer, Knowledge Management & Insights

Gates Foundation

Interface text:
1 / 26

### What’s defensible for your organization

Phipps identifies the foundation’s understanding of internal processes and tacit knowledge as its advantage. The team models both in SIP to help AI work effectively across its data.

[0:24](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=24s) [Introduction describing production constraints and internal processes and tacit knowledge as the organization’s moat.](https://assets.ai.engineer/sites/aiecode2025/sha256/01f38d933994b8f525a4aea4fca5cb369958ae5ab17befe4c16a3192cd675fca.webp)

Intro: What’s defensible for your organization

• Pace of change of AI is extremely fast but at the same time tools like Claude Code make it possible for dev to move even faster
• However, in production there are constraints:
  • how much of the deployed stack are you prepared to own?
  • how much appetite do users have for decentralized access?
  • what’s your product differentiation from SaaS platforms that are widely used?
• Forced reflection:
  • what’s your team’s skillset and competitive advantage within that environment?
• For us (and what I hope you take away from this presentation!):
  • our moat is our understanding of internal processes and the tacit knowledge needed for any AI (no matter how good!) to work effectively across our data
• These processes and tacit knowledge are modeled within the Strategic Intelligence Platform (SIP) that we launched into production on June 4 for enterprise-wide internal use at the Gates Foundation.

2 July 2026
AI Engineer World’s Fair

Interface text:
2 / 26
×  ←  →

### The Gates Foundation: Overview

The speaker introduces the foundation’s work as background for the data modeling and internal operations discussed in the talk.

[2:33](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=153s) [Section slide introducing an overview of the Gates Foundation.](https://assets.ai.engineer/sites/aiecode2025/sha256/558daef5e7100cdcce51350ee2dc0e32b1e3a8e0e8eb383a9bb7688a3d55c5b9.webp)

The Gates Foundation:
Overview

Interface text:
3 / 26

### Our role

The Gates Foundation uses data and evidence to show what works and what doesn’t. The speaker connects this role to extracting insights from grants at scale.

[2:49](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=169s) [Four columns describe market incentives, innovation, collaboration, and high-quality data, each illustrated with a photograph.](https://assets.ai.engineer/sites/aiecode2025/sha256/ad120b2c7cbe5b9efe7e432ac9e60b0ec6d13778195f2900e249d4e7f96cc72c.webp)

Our role
We work with partners worldwide to tackle critical problems

Creating market incentives for life-saving products by supporting the development and delivery of vaccines, treatments, and tools for those most in need.

Spurring innovation that improves the human condition by stepping in where governments can’t, and businesses won’t.

Strengthening collaboration by bringing together the public, private and philanthropic sectors to save and transform lives around the world.

Generating high-quality data and evidence that drives progress by showing what’s working and what isn’t.

2 July 2026
AI Engineer World’s Fair

Interface text:
4 / 26

### Scope of our work

The foundation’s 2023 snapshot spans 2,396 grants, 135 countries and 1,422 grantees, illustrating the scale of the structured data the team works across.

[3:50](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=230s) [A grid of 2023 foundation statistics includes 2,396 grants, 135 countries, 1,422 grantees, and $7.75 billion in charitable support.](https://assets.ai.engineer/sites/aiecode2025/sha256/35ee0641f1d3ab8f5b6e49d200fc937a65ae8519d5ff9519c6045269232fc1a7.webp)

Scope of our work (Stats for new grants from F.Y. 2023)
We work with partner organizations around the world to reduce inequity

2,396 Total Grants
135 Countries
2,238 Alumni
44 Program Strategies
48 U.S. States
1,422 Grantees
$7.75B Total Charitable Support
2,026 Employees

For the year ended December 31, 2023. Amount in U.S. dollars.

2 July 2026
AI Engineer World's Fair

Interface text:
5 / 26

### Foundation funding summary

Global Development and Global Health accounted for the largest shares of the foundation’s $7.75B in 2023 charitable support, at $2.005B and $1.860B. The breakdown also includes operational expenditure.

[4:37](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=277s) [A donut chart and horizontal bars break down $7.75 billion in 2023 charitable support across foundation program areas and operational expenditure.](https://assets.ai.engineer/sites/aiecode2025/sha256/47cdf39c81341b2b630ac0c437e8baf9a718ec6482114755024fddcb147f620d.webp)

Foundation funding summary

In 2023, the foundation provided US $7.75B* charitable support in the following areas:

Total charitable support
$7.75B

A. Global Development — $2.005B
B. Global Health — $1.860B
C. Global Growth & Opportunity — $875M
D. Gender Equality — $863M
E. United States Program — $692M
F. Global Policy & Advocacy — $328M
G. Other Charitable Programs — $204M
H. Operational Expenditure — $922M

*Financial figures are rounded to the nearest million and include operational expenditures, grant payments, and direct charitable support, but not program related investments (PRIs), for year ended December 31, 2023. Amounts in U.S. dollars.

2 July 2026
AI Engineer World's Fair

Interface text:
6 / 26

### Strategic Intelligence Platform

The team is building a knowledge graph that structures operational data for retrieval by agents.

[5:03](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=303s) [Section title: structuring operational data for agentic retrieval.](https://assets.ai.engineer/sites/aiecode2025/sha256/1fd4d0bc571e8de7f7092e3a11b5f12fab25ce5ec0117856f25dce777ba5cc2e.webp)

Strategic Intelligence Platform (SIP):
Structuring operational data for agentic retrieval

Interface text:
7 / 26
×  ←  →

### From Systems of Record to Agentic Intelligence

A lakehouse brings structured and unstructured data together for curation. A cross-system semantic graph lets agents reason across the data, with chat and workflows as the user interface.

[5:22](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=322s) [Diagram connecting systems of record through a data lakehouse and curation to agentic chat and workflows.](https://assets.ai.engineer/sites/aiecode2025/sha256/745575c04b0f5d030648c9916711220391da43b4cb5cd780c66fdb385975df40.webp)

From Systems of Record to Agentic Intelligence

Systems of Record
- Doc Repository
- HR Source System
- Investment Source System
- Funding Source System

Enterprise Data Platform

Data Lakehouse
- Structured Data
- Unstructured Data

Data Curation
- Pre-Processing
- Extraction & Enrichment
- Governance

Strategic Intelligence Platform
Cross-system semantic graph layer agents reason over
- Agentic Chat
- Agentic Workflows

2 July 2026
AI Engineer World's Fair

Interface text:
8 / 26

### Dataset Owner Engagement is Critical

Data owners provide tacit knowledge needed to interpret fields, join datasets and apply safeguards. They also help AI follow reporting conventions, including how questions were answered in the past.

[6:12](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=372s) [Data owners help interpret fields, join datasets, understand limitations, apply safeguards and standardize reporting.](https://assets.ai.engineer/sites/aiecode2025/sha256/c5b63dfe97a90e64c5aace837bdf0389afae91e116ca13ec7ce662149bcb66b9.webp)

Dataset Owner Engagement is Critical!

• Often significant amount of tacit knowledge controlled by data owners
• Data owner engagement critical for:
  • interpreting full meaning behind various fields
  • understanding the structure of the dataset as a whole
  • joining datasets together correctly
  • understanding limitations or systematics of the data
  • applying appropriate safeguards and security trimmings
  • ensuring standardized reporting conventions

This is the procedural understanding and tacit knowledge needed for AI to work effectively in the enterprise!

2 July 2026
AI Engineer World's Fair

Interface text:
9 / 26

### From Raw to Curated Data

Curation combines preprocessing, extraction and enrichment, and governance in dataset-specific combinations. Tags can form graph connections, while structured metadata becomes graph properties.

[7:06](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=426s) [Data curation steps grouped into preprocessing, extraction and enrichment, and governance.](https://assets.ai.engineer/sites/aiecode2025/sha256/51d2889a7a5ed937ae1e034c4f2dde774c1449a792fad5935232cf613f43d455.webp)

From Raw to Curated Data

Data Curation

PRE-PROCESSING
Filtering
Deduplication
Version Tagging

EXTRACTION & ENRICHMENT
Structured Field Extraction
Semantic Chunking
Multimodal Extraction
Entity & Taxonomy Tagging
Structured Metadata

GOVERNANCE
PII Masking
Sensitive Content Classification
Access & Entitlement Mapping

* Different datasets run different combinations of these steps.

2 July 2026
AI Engineer World's Fair

Interface text:
10 / 26

### SIP Graph Data Model: Live Walkthrough

The walkthrough uses a graph to represent an organization’s operational story and invites viewers to consider how they could model their own.

[8:21](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=501s) [Walkthrough introduction describes a graph as a practical model of an organization's operational story.](https://assets.ai.engineer/sites/aiecode2025/sha256/a27a937147c25d28be063a8bf7a4ea79334c5e7a2459435f2fa0178b513d42db.webp)

SIP Graph Data Model: Live Walkthrough

Graph is a practical representation of a physical model.
It allows us to tell the operational story that is the foundation of our business.
What's your organization's story, and how could that be modeled?

Interface text:
People — HR Source System
Investment — Grant Source System
Funding — Funding Source System
Meetings & Docs — Doc Repository
11 / 26
← → or Space to navigate · click a dot to jump

### SIP Graph Data Model: Live Walkthrough

Strategy teams’ annual reviews produce documents linked to meetings. A five-level funding hierarchy connects Division through Body of Work to funding paths and annual budgets, with IN_PATH links supporting roll-ups.

[8:52](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=532s) [A conceptual graph connects a strategy team to a meeting and its documents, alongside notes on annual planning.](https://assets.ai.engineer/sites/aiecode2025/sha256/0d085a4b7b3e70810d5b2cd47c023e4064b8a85bd9123352c6a4a72bd5984ebf.webp)

STEP 1 · THE ANNUAL RHYTHM
Strategy reviews set the agenda

• Each year strategy teams review progress and set priorities across a 5-year horizon.
• Strategy Reviews and Annual Planning generate board decks, rankings and notes.
• Those documents live in the curated document library — our unstructured corpus.
• 80+ strategy teams across the foundation.
• What's shown here is the conceptual data model, which doesn't show instantiation or cardinality.

Diagram:
Strategy — strategy team
HAS_MEETING
Meeting — Strategy Review · Annual Planning
HAS_DOCUMENT
Document — board decks · narratives

Interface text:
INVESTMENT
FUNDING
MEETINGS & DOCS
People — HR Source System
Investment — Grant Source System
Funding — Funding Source System
Meetings & Docs — Doc Repository
12 / 26
← → or Space to navigate · click a dot to jump

[9:31](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=571s) [A five-level funding hierarchy connects Division through Body of Work to a funding path and annual budget.](https://assets.ai.engineer/sites/aiecode2025/sha256/884f4fa9bad13f859ab73114cbf2f8f11e8a611d9eb2e3156093cae25e0c832e.webp)

STEP 2 · FUNDING THE PLAN
Budgets flow through the funding hierarchy

• Annual planning finalizes each strategy team’s board-approved budget.
• Money flows from a strategy to a funding team defined by a 5-level funding path: Division → Strategy → Group → SubGroup → Body of Work.
• A funding team is allocated an annual budget (FundsToBoW).
• Aliases and full text indexes on team nodes to enhance agentic retrieval.
• Funding path is a 5 level additive DAG.

Diagram:
Division — BoW hierarchy
CONTAINS
Strategy — strategy team
CONTAINS
Group
CONTAINS
SubGroup
CONTAINS
Body of Work — funding entity
IN_PATH
Funding Path — BoWHierarchyPath
MANAGES_BOW
FundsToBoW — budget in · yr
HAS_FUNDS_TO_BOW
Meeting — Strategy Review · Annual Planning
Document — board decks · narratives

Interface text:
FUNDING
MEETINGS & DOCS
People — HR Source System
Investment — Grant Source System
Funding — Funding Source System
Meetings & Docs — Doc Repository
13 / 26
← → or Space to navigate · click a dot to jump

### Funding paths become investment portfolios

Funding paths link a body of work to investments and record committed or paid funding by year. Multiple funding teams can fund one investment; a body of work’s portfolio combines the investments its funding paths reach.

[10:21](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=621s) [A funding path connects a body of work to investments and annual committed or paid funding.](https://assets.ai.engineer/sites/aiecode2025/sha256/e083716227e8858b0bae888b25a6c319f12fe45ae67b63258e8107d32ffd662f.webp)

STEP 3 · PORTFOLIOS FUND INVESTMENTS
Funding paths become investment portfolios

• Each funding path commits and pays out to specific investments (FundsToInv, by year).
• A body of work’s investments are the union of everything its funding paths reach.
• Here the green funding world hands off to the yellow Grant Source System world.

Diagram:
Investment — multi-year engagement
Body of Work — funding entity
Funding Path — BoWHierarchyPath
FundsToInv — committed/paid · yr
FUNDS_INV
HAS_INV
MANAGES_BOW
HAS_FUNDS_TO_INV

Faded diagram nodes:
SubGroup
FundsToBoW — budget in · yr

Interface text:
People — HR Source System
Investment — Grant Source System
Funding — Funding Source System
Meetings & Docs — Doc Repository
14 / 26
← → or Space to navigate · click a dot to jump

### Three hierarchies, aligned

RELATED_TO links align divisions and connect WorkTeam, Team L2 and Strategy across three source systems, allowing an agent to traverse the organizational structure.

[13:09](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=789s) [Related-to links align division and team nodes across the people, investment and funding hierarchies.](https://assets.ai.engineer/sites/aiecode2025/sha256/85f1f26eaafa135a9eaf1e313f1255763891ab9f3b68f0740e8081eff97554b2.webp)

STEP 7 · THE CROSS-SYSTEM SPINE
Three hierarchies, aligned — the moat

• People, Investment and Body-of-Work hierarchies are stitched together by RELATED_TO.
• Division ↔ Division ↔ Division, and WorkTeam ↔ Team L2 ↔ Strategy — three systems, one backbone.
• This semantic alignment is the durable work no model or framework hands you for free.

PEOPLE
HR SOURCE SYSTEM
Division
COMING
People hierarchy
CONTAINS
WorkTeam
COMING
Employee
owner · director · attendee

INVESTMENT
GRANT SOURCE SYSTEM
Division
Investment hierarchy
CONTAINS
Team · L2
Team · L3
Team · L4
manages investment

FUNDING
FUNDING SOURCE SYSTEM
Division
BoW hierarchy
CONTAINS
Strategy
strategy team
Group
[illegible]

RELATED_TO
RELATED_TO
RELATED_TO
RELATED_TO

Interface text:
People — HR Source System
Investment — Grant Source System
Funding — Funding Source System
Meetings & Docs — Doc Repository
18 / 26
← → or Space to navigate · click a dot to jump

### Documents become discoverable in context

The model links meetings to documents and their AI-tagged content segments, then maps segments to structured entities. A full-text index can aid search and retrieval.

[13:47](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=827s) [Graph highlights meetings, documents and content segments linked to structured entities.](https://assets.ai.engineer/sites/aiecode2025/sha256/1aef77a896175f4b3da9ec8033c9a39b584ec12921dc1ee4566c819688f1dcc2.webp)

STEP 8 · THE UNSTRUCTURED OVERLAY
Documents become discoverable, in context

• Documents are semantically chunked into ContentSegments — tagged and summarized by AI at ingestion.
• Multimodal extraction lifts both structured data and unstructured narrative out of every document — tables, charts and prose alike.
• Segments map to the structured entities they discuss, so narrative sits where the data lives.

Graph nodes: Division; Team · L2; Team · L3; Team · L4; Strategy; Group; SubGroup; FundsToBoW; Funding Path; Body of Work; FundsToInv; Meeting; Acronym; Attendee; Document; FollowUpItem; ContentSegment; Top5Bottom5

Graph relationships: USES_ACRONYM; HAS_ATTENDEE; HAS_DOCUMENT; HAS_FOLLOW_UP_ITEM; SOURCED_FROM; ASSESSMENT; HAS_SEGMENT; HAS_SUBSEGMENT; ASSESSES; MAPS_TO

Graph captions: strategy team; Strategy Review · Annual Planning; role at meeting time; board decks · narratives; AI-chunked · tagged; full-text index; claimed rankings; funding entity; [illegible]

Interface text:
People — HR Source System
Investment — Grant Source System
Funding — Funding Source System
Meetings & Docs — Doc Repository
19 / 26
← → or Space to navigate · click a dot to jump

### UX with SIP

Claude checks the SIP schema and runs a Cypher query to explain 1KD and identify its source document. Enterprise workflows are still being built.

[15:24](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=924s) [Sample Claude chat querying SIP beside an enterprise workflows placeholder.](https://assets.ai.engineer/sites/aiecode2025/sha256/1ded40c194a75d1f2dc01c7419ba008f63aa3a0f4014e167750cecd26d599e49.webp)

UX with SIP
Agentic application interfaces

Chat: Claude + SIP

What does 1KD mean in SIP, and which document defines it?

I need to look into the SIP (Strategic Intelligence Platform) to answer this. Let me check the schema first.

Loaded tools, used SIP: Strategic Intelligence Platform integration
Finding tools
Result
Get Neo4j Schema
Result
Done

“1KD” is an acronym defined in a specific meeting document. Let me look it up.

Read Neo4j Cypher
Read Neo4j Cypher
Result
Done

In SIP, 1KD stands for “First 1,000 Days” (also written “First 1000 Days”) — referring to the period from conception through a child's second birthday, a common framing in maternal and early-childhood health work.

It's defined in the document “2025 10 14 Neuro Learning Session Acronym List.docx”, from the October 14, 2025 Neuro Learning Session.

Sample query of SIP

Enterprise Workflows
Coming soon ...

2 July 2026
AI Engineer World's Fair

Interface text:
21 / 26

### Process for generating structured evals

Data owners and the eval team developed 93 reporting questions across complexity tiers. Each question has a reference query rerun at eval time, so its answer can change with the data.

[17:20](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1040s) [Data owners and an evaluation team develop reporting questions paired with re-executable reference queries.](https://assets.ai.engineer/sites/aiecode2025/sha256/c368a8985cdec7889f15f115a92ac36d4cc00618aa977d1b83a8014a1b36e888.webp)

Process for Generating Structured Evals

• Evals allow us to understand gaps/ambiguities in our data model or the ways in which users interact with it
• Enable us to ensure performance stability over time as the data is refreshed or data model is updated

HUMAN COLLABORATION

Data Owners
(subject-matter experts)
know the graph data + reporting conventions

Workshop
Data owners + eval team surface standard reporting questions

TECHNICAL ARTIFACT

93 Standard Reporting Questions
Complex analytical reasoning
Multi-hop / aggregation
Simple lookup questions
Simple
complex

CHALLENGE
The correct answer changes with every data refresh — there is no static ground truth.

SOLUTION: pair each question with a re-executable reference query
Question
Question
Reference Cypher query (re-run at eval time)
[illegible] live reference answer, never frozen

2 July 2026
AI Engineer World's Fair

Interface text:
22 / 26

### Structured eval pipeline

An LLM judge compares the agent’s answer with a reference answer from the current knowledge graph. Pass@1 and stability results guide changes to the data model, schema descriptions and domain rules before the pipeline is rerun.

[18:19](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1099s) [Reference and agent answers feed an LLM judge; results guide data model refinements and reruns.](https://assets.ai.engineer/sites/aiecode2025/sha256/6573c14d19739ac85db1b47d4fc1f836b1f2578fc4e97eca6d0a16709a1e9440.webp)

Structured Eval Pipeline
Claude Agent SDK Batch Pipeline

Eval Question (from eval set)

REFERENCE PATH
Run reference Cypher on SIP Knowledge Graph (current snapshot)
Reference Answer

AGENT PATH
Claude + SIP
tools: Cypher + MCP
Agent Answer

binary: is agent answer consistent with reference?
LLM JUDGE
Consistency Metric

pass@1
single-run binary score

Stability (k=3)
run 3x, modal agreement

Eval Results

Failures + low stability reveal ambiguities and data modeling gaps
Update data model, schema descriptions, domain rules
Re-run pipeline

FEEDBACK LOOP: results -> refine data model -> re-run eval

2 July 2026
AI Engineer World's Fair

Interface text:
23 / 26

### SIP’s structured eval results

Across 93 questions, pass@1 rises from 89.3% with Sonnet 4.6 Low to 94.6% in the other configurations. Stability ranges from 96.0% to 98.9% (k=3). The speaker says missed questions tend to be ambiguous.

[18:56](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1136s) [Bar chart compares pass at one and stability across four model and reasoning configurations.](https://assets.ai.engineer/sites/aiecode2025/sha256/9a1dc474c27a5a1b654e0c1e1e3cfaef67def84d94b8f24ee8821679fdf9ba07.webp)

SIP’s Structured Eval Result Summary
Model + reasoning dependent performance and stability

Overall Performance — Pass@1 vs Stability (n=93 questions, k=3)
Y-axis: Score (%)
Y-axis ticks: 75, 80, 85, 90, 95, 100
Sonnet 4.6 Low: Pass@1 89.3%; Stability 96.0%
Sonnet 4.6 Medium: Pass@1 94.6%; Stability 97.9%
Sonnet 4.6 High: Pass@1 94.6%; Stability 98.2%
Opus 4.8 Medium: Pass@1 94.6%; Stability 98.9%
Legend: Pass@1 (solid); Stability (hatched)

2 July 2026
AI Engineer World’s Fair

Interface text: 24 / 26

### What’s ahead for SIP

SIP plans to expand its enterprise graph, link team-specific data through federation, and build agentic workflows for controlled, auditable outputs.

[19:24](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1164s) [Roadmap for expanding SIP data and graphs and building governed agentic workflows.](https://assets.ai.engineer/sites/aiecode2025/sha256/2debe4e264b61518203cb8730a8e5a459b2f969e1b7b20c636f9c5fe4aa15d1a.webp)

What’s ahead for SIP

• Continuing to fill out data from existing systems of record (i.e. fits cleanly in our current data model)
• Controlled expansion of primary graph to additional enterprise-wide datasets (i.e. data model expansion)
• Federated graph experience to link team specific datasets to enterprise-wide subgraphs
• Build agentic workflow experiences that allow for controlled generation of artifacts, reports, search & retrieval, etc
  • Compared to general chat: enables standardization, optimization, and auditability (i.e. more controlled, governed experiences)

2 July 2026
AI Engineer World’s Fair

Interface text:
25 / 26

## Full transcript

[0:00](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=0s) [upbeat music] Yes, my talk today is about the title, Your Data Model is Your Moat.

[0:16](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=16s) We have a enterprise-wide platform that we had just rolled out here this past month, and so I'll go into details on this. I'll give you some, hopefully some practical lessons here and why we made decisions we made for this, uh, how, how you could picture your processes within a similar type framework.

[0:35](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=35s) So first, just a, a quick introduction. So this gets into the, the title here of the talk, and the, the framing of, you know, what I hope you take from this.

[0:43](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=43s) But with AI moving very fast at the frontier, what, what's defensible? You know, you can move, you can build things very quickly with Clo- Claude Code, um, but once you push things to production, there's constraints.

[0:59](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=59s) You know, you find how much of your, your, uh, deployed stack do you wanna actually own? You know, there's monitoring, there's upkeep, there's, uh, you know, people building dependencies off your stack that you have to be prepared to handle.

[1:11](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=71s) How much, uh, appetite do users have for decentralized access? This, this gets into... Uh, I'll show you what, what we built, but, you know, what, what's, what's the access point for users?

[1:21](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=81s) You know, is it another chat app? Is it Claude? Is it ChatGPT? Uh, is it something else? What's your product differentiation from, from those different SaaS products? And so our team then, you know, with this context in mind, you know, thought through here, you know, what, what's our skill set here?

[1:39](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=99s) What's our competitive advantage in this environment? And

[1:43](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=103s) this is what I, you know, really hope that you, you take from this talk and, you know, picture yourself in this. But our moat here was our understanding of our internal processes, the tacit knowledge that you need to, to run successful AI.

[1:58](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=118s) And, and this is true, I think, no matter how good AI gets, how good models get, you know, new releases that different companies put out when, when, uh, Mythos comes out or when there's a new app from Claude.

[2:08](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=128s) Yeah, I'm not, I'm not worried because the part that we've built is the defensible part that, that's, that's durable.

[2:16](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=136s) So these are the... I'll, I'll, I'll tell you what this means here in, in, in more detail, but these are the, the processes, tacit knowledge that we've modeled into what we call the Strategic Intelligence Platform, or SIP.

[2:26](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=146s) And it rolled out here this past month in production for enterprise use across, uh, f- the Gates Foundation, so about four thousand people.

[2:36](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=156s) So first, uh, I know this is an engineering talk, but the, the, the scope of this talk gets into data modeling, internal operations processes, and so I wanna give very quick background here over what the Gates Foundation does because then this is what we're, we're modeling.

[2:52](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=172s) So as you're probably familiar, the, the Gates Foundation has a, has a very wide scope and it's a very ambitious work that we've been doing for the past twenty-five plus years.

[3:02](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=182s) And there's all kinds of, you know, broad initiatives that we're doing, you know, whether it's for, uh, child mortality, whether it's for nutrition, agriculture, uh, education. And these are kind of broadly the, the different buckets that these different initiatives fit, fit into.

[3:19](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=199s) Creating market incentives, spurring innovation, collaboration between public and private sectors, and then the fourth one here kinda gets into the, the lens that we're building here. You know, high-quality data, trying to derive data-driven insights from the, from the actual investments, the, the grants that we've put out.

[3:39](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=219s) And over twenty-five years, there's a ton of structure, there's a ton of data that's developed, and trying to extract those insights at scale i,i- is difficult, and that's what we're trying to solve.

[3:52](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=232s) So this slide here is a snapshot of the, of some of the different, uh, of the work that went out in twenty twenty-three within the foundation. Th- this gives you an idea.

[4:02](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=242s) I just put this here to, to show some of the structured, the, the structure that we have that we're, we're working across. So you have over two thousand grants in one year.

[4:11](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=251s) So ma- many of these are five million plus. Uh, many cou- hundred-plus countries that, uh, that, that are targeted with these grants. Uh, alumni, so there's four thousand different employees of the foundation.

[4:24](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=264s) Um, you know, many different strategies within the foundation. The US, within the US, across al- almost all the states. Grantees. The total annual disbursement, over seven billion dollars.

[4:37](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=277s) And so this gives you some idea of structure that we're, we're working with. And, uh, this one, just finally here, when I show the data model, this will make more sense, but we have different divisions.

[4:46](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=286s) That, that funding goes out through different divisions. And so this breaks down some of those divisions so you can see different priorities, and it'll, it'll make more sense in a second here.

[4:56](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=296s) But global development, global health, uh, gender equality, USP are just a sample of the different divisions.

[5:05](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=305s) Okay, so the, the fun stuff here now, I hope. The, uh, Strategic Intelligence Platform, so

[5:12](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=312s) i- in a, in a, in a nutshell here, structuring operational data for agentic retrieval. So we're building a knowledge graph with the idea of the agent consumer.

[5:24](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=324s) And here is an end-to-end look of what this looks like. So we, we have different systems of record, structured, unstructured. These have been siloed traditionally.

[5:36](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=336s) The... So part of our team here, the work has been to create what's essentially a da- uh, data lakehouse, putting everything under one roof. This is our internal enterprise-wide data.

[5:47](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=347s) It's also different, different programmatic data that are, uh, uh, outputs of different investments. Once it's there, it's easy for us to consume, so we have a data curation layer that does different processing to it.

[5:59](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=359s) And then finally, SIP here at the end with, uh, agentic chat, agentic workflow as the-

[6:04](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=364s) As the, the U- the UX, how users are consuming our platform. And so it's a cross-system semantic graph layer that agents can reason, reason across.

[6:14](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=374s) Okay, uh, so some of this I'll try to speed through here just for the sake of time, but th- this one is critical.

[6:20](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=380s) When you're dealing with systems of record with lots of complexity, engagement is critical. This is something that we've, we've found here repeatedly. We have to engage data owners to understand, you know, this tacit knowledge we're trying to, to model.

[6:33](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=393s) What's the full meaning of different fields, the structure of the dataset? How do we join things together? How do we, uh, understand limitations, systematics of the data, safeguards, security trimmings, uh, reporting conventions?

[6:45](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=405s) You know, it's not enough just to answer a question a certain way. You have to answer it the way that it's been answered in the past. And so this is the-- comes back to the moat here.

[6:55](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=415s) This is the procedural understanding, tacit knowledge that AI needs, and it's, you know, it's the part that we, that we own, that's, you know, that's ours. That, um, and that's what we're, we're modeling here.

[7:09](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=429s) Okay. So going back here just very quickly for this one, this is the, a snapshot here of different data curation considerations that we're-- that go into this pipeline. So you have, for different datasets, whether it's structured, unstructured, there's different preprocessing, filtering, deduplication.

[7:24](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=444s) There's an order to different documents. There can be, uh, inconsistencies across documents. Those need to be, uh, handled up front. There's extraction, so structured field extraction, semantic chunking for unstructured documents.

[7:36](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=456s) If you have figures, you need to convert this into text in some way so you can do retrieval across this. Uh, various forms of tagging that th- these can form connections in your graph.

[7:46](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=466s) Structured metadata that you create during this pipeline, and that becomes different properties in your graph. And then the third bucket here, governance, and this is a, an important one that, uh, I think AI makes more acute.

[7:59](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=479s) Things that were, that were accessible previously, they're much more accessible now with, with AI, and so you have to consider this. Your, your risk sphere is, is larger. So things like PII need to be masked.

[8:12](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=492s) You need to reconsider different, uh, sensitive data, classifying this, um, making sure that, that there's the right entitlements for each user who's accessing your system.

[8:23](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=503s) Okay, so that's the overview here. The, this, the, the data model itself. Now, this is the part I'll walk through here.

[8:30](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=510s) There's a, a nice animation here, but hopefully the takeaway is you can picture your own, your own organization's story within what I show here. I'll get somewhat technical, but it's only to hope- hopefully to give you an idea of how we, how we solved our problem, and then you can hopefully, uh, model this to yours as well.

[8:48](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=528s) Graph is a very flexible, um, practical representation of a physical model.

[8:54](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=534s) Okay, so I'll zoom through a few of these here, but the-- just the, the entry point here, we have over eighty different strategy teams. These teams have annual reviews that happen.

[9:04](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=544s) This is how the budgeting for each year is derived. And then so we model this here in the graph. The, the, the meetings are where unstructured documents, uh, enter into this system from, but then there's-- they have a structured connection to your other systems of record.

[9:19](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=559s) Um, what I show here is a conceptual data model, so it's flat. So you're not seeing instantiation. You have the actual graph. There's many different nodes. Cardinality, is it one-to-one, one-to-N?

[9:29](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=569s) So the actual graph, it's, you know, even more complicated. But

[9:34](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=574s) for the data model itself, let's-- let me, let me show you the first different hi-- So we have multiple hierarchies that exist within what we've modeled. The-- There's different types of hierarchies you can have.

[9:43](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=583s) In this case, this is a, hopefully you can see all this very well, but it's a, it's an, it's a, um, additive DAG. So there's a-- all five levels here of this hierarchy from the top to the bottom matter.

[9:59](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=599s) So you have to consider everything together. And so then there's different roll-up patterns you can do to work across this, this sort of pattern. In our case, we have a in path shortcut here that connects the funding path.

[10:12](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=612s) Uh, funds to BAU is where we have the, um, the budget for each of these different funding teams that, that's stored.

[10:24](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=624s) So we have funding. What's the, the internal funding teams have portfolios. These portfolios then go towards different investments. Multiple funding teams fund an ind- individual investment, so it's a N-to-N relationship there.

[10:38](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=638s) The investments are the thing that are our product. It's our, it's our, our business. But internally, we have funds that then prioritize different, different types of investments, and that's what's shown here.

[10:50](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=650s) And so you can take this down to the transaction level, or you can have different, uh, annual-based aggregations that you map here as well.

[10:59](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=659s) And then from investment, there's a lot of interesting things you can do. You can map to all the different organizations, and you can have different types of organizations, and there's actually a lot here that is still kind of a green space that we wanna fill in.

[11:11](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=671s) We have all these different observables that people have produced in the investments that we wanna model here. So published reports, products, you know, all this stuff is structured and connects to the entire, uh, organizational picture.

[11:27](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=687s) So I mentioned that there's different hierarchies. This is the second type of hierarchy. At this hierarchy, each level matters in and of itself, and so it's not a, a DAG necessarily, and so you can actually do things like pre-computing the, the-- some of these, these different shortcuts.

[11:44](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=704s) So the hierarchy, it goes from the top to the bottom, contains, connects it. This is showing the investment management side of the, of the organization. And there's concepts of direct team management, so one team at, like, team level two manages the investment.

[12:01](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=721s) But then there's also a concept of indirect management, so that the, uh, children below team level two still should be attributed to the team level two. And so there's different things you can, different games you can play with these sort of roll-ups to pre-compute.

[12:15](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=735s) I don't know if you can see this, but roll-up Manages Inv is a, is a, a derived edge that we, that we create after we create the Contains and Manages, Manages Inv edge.

[12:28](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=748s) So I've shown two different, two different lenses for one investment. There's the funding lens, the management lens, and you can model both of these here then within, within the graph.

[12:38](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=758s) A third hierarchy here is people. You have organizations, you have org charts, and you have people who are owners, you have people who are attendees in meetings, you have people who are, uh, directors.

[12:48](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=768s) Th- there's all kinds of different roles they have. You can model these here. You can have their, their, uh, you know, who they report to, what their, uh, team structure is.

[12:57](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=777s) And th- these all are structured data that connects across systems. Traditionally, they existed in a, just a HR source system, but they're relevant for the context of the, the full story.

[13:11](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=791s) And then that leads to this connectedness. So we have different source systems that were siloed. We-- To understand the entire picture, for the agent to understand correctly across the structure, you need to find these common ba- these common, uh, these common, uh,

[13:29](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=809s) entities that you stitch together. And so that's what's shown here. These are different source systems, but they're related quantity, entities that exist there. And now the agent can traverse here and understand this pretty complicated organ- organizational structure.

[13:49](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=829s) One last part here that I haven't shown yet is the, the document part. So... And this is still, there's, there's a lot more we can do to, to this part.

[13:56](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=836s) We've just been, uh, ingesting one different document source so far. But this is where you combine unstructured and structured, and this gets into part of the magic here that you can model with Neo, Neo4j.

[14:08](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=848s) But we have meetings that have documents, documents that have different semantic sections that you can, or chunks that you can, uh, that you can model here. You can put full text indexes across these to, to aid in the different, uh, search and retrieval approaches for the agent.

[14:26](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=866s) There could also just be a pure graph retrieval that, that the agent does. And then all these things then connect back to your, your, your main organizational structure.

[14:38](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=878s) So then as a whole, th- this is what the data model looks like. So I've been zooming in here. Now you can see the, the full interconnectedness of this.

[14:45](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=885s) Four different systems, one graph, uh, one semantic layer that's exposed through an MCP then to the, to the agents.

[14:56](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=896s) And so this is the-- So if you think of the agent's perspective, this is the, the structure that it can dynamically discover and reason across at query time. And for the developer, it's also a very cool thing because it exposes, you know, what you don't know about your, your, the thing you're modeling.

[15:12](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=912s) You know, you very, very soon you find out that there's a gap in your understanding or there's some data set that you're not, you know, fully including. And so this, this process in and of its, in and of itself is very valuable.

[15:26](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=926s) Okay. Let me give you a sense here what we do with this now. So this is the-- I showed you the platform, the graph, but then how does this relate to AI?

[15:34](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=934s) So we've, we've connected this through MCP, and I, I, you know, I discussed earlier what the, what's durable, what's defensible. To us, what was not defensible was the, was the, the chat interface, was the UI, and even in some cases, the, the general chat cases, the, um, you know, the, the agent interaction.

[15:54](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=954s) And so we, u- users themselves are in Claude already or ChatGPT, and so we serve the platform where they are. And so it's served here now through MCP. Here's an example, just kind of a innocuous, uh, question here.

[16:09](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=969s) But Neo4j has some off-the-shelf, uh, MCP servers here. We, we've actually modified these quite a bit here. We forked it, and then there's various updates to the schema, uh, things to, to pass state back to the, to, to our system.

[16:25](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=985s) You know, the, the conversation, uh, IDs, the, the message, uh, numbers, stuff like this we, we, we've, we've modified in these MCP tools.

[16:35](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=995s) But so that's a general chat experience. That's one entry point. The other part that we're building right now too, that's very exciting, is more constrained workflow experiences. And so these can also be offered through things like Cowork or, uh, Claude Chat.

[16:48](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1008s) And you can do things like, um, you can have your, you can have MCP apps be the, the, you know, the standard entryway that users access. So, you know, different UIs that are, uh, ported into your, your, your, your chat experience.

[17:02](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1022s) And you can have different sandbox-based agents that then run the, the workflow. And so these are, these are active things that we're working on. It helps to constrain the experience compared to chat, but it, at the same time, it pulls from that same knowledge graph-based, uh, backend platform.

[17:23](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1043s) Okay. I've got a couple of minutes. I'll, kind of speed through this. But the, the, the way evals then relate to data modeling is that a- as you're doing evals, you find, you find gaps, you find ambiguities in your data model.

[17:36](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1056s) You find ways in which users are asking questions that, uh, that are ambiguous or it's, um, you know, not, it's not returning things that conform with the reporting standards.

[17:47](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1067s) So what, what we've done here then is we've worked with data owners. We've d- we've, uh, built targeted, uh, eval questions that, that they, that match their reporting standards.

[17:59](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1079s) We've separated these into different complexity tiers. One challenge is that the, the structured data is constantly changing, so we have to have the graph query itself that we, that we create for each of these different questions.

[18:11](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1091s) And then at runtime for the evals, we, we pull from the live graph, and then we compare that to what the agent is delivering for that question.

[18:20](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1100s) And so that's what's shown here then. And there's a feedback loop that you can do for this. So as you're running, uh, an eval pipeline, an eval structure pipeline, you have an LLM as a judge.

[18:28](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1108s) We've modeled things like pass@1, uh, stability. So if you ask the same question multiple times, you get the same answer back. You can use LLM-as-a-judge to, to, to measure this.

[18:39](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1119s) And then there's a feedback loop here that you, you can update then your, your data model, you can update your, uh, your domain rules, your, uh, schema descriptions to help, to help, uh, fill those, tho- those gaps that you, that you find.

[18:58](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1138s) Then after you do this, th- this is, uh, this is just some, some, some eval reporting here that we, uh, we show the pass@1 and the, the stability for our system.

[19:08](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1148s) So we- we've gotten this very, ver- very strong. The, the questions that we, that we end up do missing, it tends to be things that are ambiguous in some way.

[19:15](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1155s) And so it's not wrong, it's just that it's things that

[19:20](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1160s) might be right, but not what the user intended. So that, that's kinda the constant struggle that we, that we, that we're working around.

[19:27](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1167s) 30 seconds here. What's ahead for SIP? So we're continuing, continuing to fill out our existing, uh, data from system, systems of record, so things that fit into our current data model.

[19:36](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1176s) We wanna expand the primary graph to additional enterprise-wide datasets. There's a lot of, there's a lot of demand for a federated graph experience. So we have a, a main enterprise system, but we have specific teams that have their own data that they wanna link to this, and so we're working on how to do this.

[19:53](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1193s) Uh, different agentic experiences, like I mentioned as well.

[19:58](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1198s) And that's it. Uh, so yeah, please, if you wanna ask questions, if there's things that you wanna talk about, I'll be out back, or you can add me on LinkedIn here and, you know, keep the conversation going.

[20:08](https://www.youtube.com/watch?v=jt1Pbr_n6oU&t=1208s) Thank you. [audience applauding] [upbeat music]
