Karthic Rao

Karthic Rao is a software engineer and the creator of Incredible, a tool for making developer videos. His work spans object-storage infrastructure, developer education, and AI workflow systems. In his writing at Perpetual, he examines a problem that ordinary information retrieval leaves unresolved: how an agent should track changing commitments and act on them without losing the context, permissions, or approvals that make an action legitimate.

From open source to developer products

Karthic grew up in India’s Western Ghats and entered software through Free Software Movement Karnataka. The community introduced him to Linux systems programming, device drivers, technical blogging, and teaching. Contributions to the Caddy web server opened professional opportunities, leading him to MinIO, where he worked on object storage.

He subsequently moved into consulting to broaden his responsibilities beyond engineering. Working with startups brought sales, product decisions, and customer expectations alongside implementation. In April 2019, he joined Dgraph Labs to manage product and developer marketing. He wanted to help shape software adoption as well as build software, and graph databases interested him for their ability to represent connected data beyond rows and columns.

With Incredible, Karthic made developer communication a product. Its workflow combines Markdown-supported storyboarding, collaborative editing, screen recording, and video production. Examples include Rust tutorials and a Replit integration that lets viewers run code alongside a video. The project brings preparing an explanation, demonstrating software, and letting a viewer experiment into the same workflow.

His later work at MinIO returned to infrastructure, particularly Apache Iceberg and the systems that make object storage usable as a transactional data lake. In his case for integrating Iceberg catalogs into object storage, he identifies a practical security problem: permissions for discovering and changing a table can live in a different system from permissions for its underlying files. He advocates bringing catalog operations and storage governance together so administrators can secure the logical table without coordinating disconnected permission models.

Memory that preserves changing commitments

Karthic’s writing at Perpetual addresses agents that remember conversations, coordinate people, and carry work across days. Retrieving a relevant passage is only part of that task. An agent must also distinguish a proposal from an accepted commitment, remember who owns it, and recognize when later conversation changes the plan.

• Selective memory indexing. Karthic argues for designing memory around the queries an agent must answer. A question about the week after a vacation requires resolving dates and filtering events; semantic similarity alone cannot do that work. He uses Chronos, a research system developed by others, to illustrate how a compact time index can direct an agent back to preserved dialogue. The design choice is what to extract in advance while retaining the original context for later interpretation.

• Version-controlled operational memory. He treats an agent’s understanding of owners, deadlines, dependencies, and permissions as state that needs reviewable changes. If a conversation reassigns an email from Alex to Morgan, overwriting a task record loses the reason for the change. His proposed approach preserves an event history, supports comparisons between states, and allows mistaken interpretations to be reversed. Competing interpretations from multiple agents require reconciliation rather than allowing the last write to win.

• Interpreting commitments. His critique of action-item extraction centers on how people negotiate obligations. A request may be conditional, spread across several exchanges, or assigned through a third party. Turning those exchanges into JSON does not settle who agreed to do what. He argues that systems should seek human validation or abstain when the commitment remains unclear.

Turning a request into an approved action

Karthic’s approach to enterprise automation begins with bounded actions that have defined inputs, permissions, success criteria, and an audit trail. Human corrections and approvals become records from which a system can learn how the organization works, supporting gradual delegation.

In his presentation on workflows for finance, wealth, and insurance, he makes the difficulty concrete: filling out a form for a client can require identifying the right household, gathering missing information, contacting the client, waiting, following up, obtaining a signature, and retaining evidence. A fixed flowchart struggles with ambiguity; an improvising agent introduces uncertainty about which actions it will take. The design he describes at Perpetual assigns control flow and external actions to a deterministic system, while models interpret information and make judgments at explicit, typed points.

A planner translates a request into a graph whose nodes refer back to the request. A compiler checks the proposed plan against the firm’s action contracts and can return a valid plan, ask for clarification, or refuse. Approved plans are immutable, and permission to perform an external action is tied to its specific payload. Waiting and followups are explicit state-machine steps. If an action’s outcome is uncertain, the system sends it for manual review rather than automatically retrying and risking a duplicate action. Models can propose smaller plans as new information arrives, but those plans face the same checks and approvals.

Karthic also describes using typed, logged judgments to support smaller models and, eventually, privately hosted inference and training. His illustrative allocation of 19 out of 20 decisions to a smaller model is a design target, not a measured deployment result; the training loop remains forthcoming. The proposal extends his work on operational memory: preserve what the system understood, what changed, and what people approved before using that history to delegate more.