The Factory That Dreams: 39 AI Agents, No Framework

Who is this customer? What did we quote them in 2019? Why did their machine need that unusual modification? At Machinecraft, those answers were part of the company’s most valuable infrastructure, but they lived largely in people’s heads. Rushabh Doshi describes running a 100-person factory in India without a data science team or an ML budget. Its institutional memory had passed from his grandfather to his father and then to him.

Employee turnover made that arrangement increasingly fragile. Every departure took knowledge with it. The fear was less about competitors than about waking up to discover that the business depended on two increasingly tired people remembering everything.

Writing another document would not necessarily solve the problem: someone still had to find it, read it, and apply it. Doshi wanted an active memory—a company twin that could use the accumulated history in everyday work. His first practical expression of that ambition was to build a sales team in software.

Machinecraft makes thermoforming machines: equipment that heats a plastic sheet and shapes it. The same core machinery can produce hydroponic farm trays, spa bathtubs, EV panels, medical casings, or packaging. Doshi describes seven buyer worlds; these are the five applications he names. A useful sales system therefore needs more than brochure recall. It must recognize which industry a customer belongs to and interpret the machine’s capabilities in that customer’s context.

The starting material was the company’s own history: years of quotations, drawings, payment schedules, timelines, and email threads. Doshi describes hundreds of gigabytes of private records. This was the relevant information environment—not the public web, but the records of what Machinecraft had actually promised, designed, and delivered.

There was no model training or fine-tuning. Instead, the ingestion process used off-the-shelf models to organize existing information:

1. Split the records into manageable chunks.
2. Have models read those chunks and extract facts.
3. Store representations of chunk meaning as vectors.
4. Store connections between entities as graph relationships.

The selected slide names Qdrant for vectors and Neo4j for relationships. These stores serve different purposes: semantic representations help locate relevant material, while graph connections preserve who or what is related. The central investment is organized memory, rather than a more capable trained model.

That memory became part of Ira, a system Doshi describes as something the company was raising. Biology supplied a vocabulary for separating responsibilities:

The purpose of the metaphor is coherence over time. Ingestion, retention, and protection against bad information are distinct jobs, even though they all contribute to the same company memory.

Doshi describes 36 agents in the spoken walkthrough. Their purpose is to divide work that would otherwise compete inside one all-purpose prompt. A pricing decision, a machine specification, and a factual correction require different responsibilities and checks.

For the correction guardian, Doshi sets an ambitious requirement: once a human fixes something, it should stay fixed permanently. That is the intended behavior, not a measured durability result.

The agents also collaborate. Athena convenes the relevant specialists, their contributions can disagree, and the process produces one answer. This combines narrow ownership with coordination: each agent has a bounded job, but the final response can incorporate several kinds of expertise.

Ira’s operating scope is the commercial front of the business: the work between discovering a potential customer and converting that prospect. Doshi reports nine daily jobs; the spoken examples identify seven:

• Outbound email: Draft outreach grounded in Machinecraft’s actual business context.
• Account briefs: Assemble cross-checked information before a call.
• Quotations: Prepare commercial quotes.
• Outreach selection: Offer a swipe-left, swipe-right review mode.
• Lead revival: Revisit dormant prospects through a workflow Doshi calls “blast from the past.”
• Inbound replies: Prepare responses to incoming messages.
• Fit screening: Determine whether a company is a plausible customer before spending an hour on it.

These workflows give the stored history a practical destination: information becomes preparation, qualification, or a proposed commercial action.

The operator works from one Cursor tab. A request can trigger knowledge-base searches, inbox reading, email drafting, and quote construction. Ira then presents the work before anything goes out. The interface is small; the supporting system is not.

Behind it are separate databases for vectors, the relationship graph, and CRM records. Doshi reports using three model providers, selected according to the task. Google integrations, document ingestion, communication tools, and monitoring connect the agents to the business and make their activity inspectable.

Doshi reports 213 tools exposed through a single protocol. The governing rule is Ira drafts; a human sends. Tool breadth does not remove the approval boundary: preparing a message or quotation is distinct from authorizing it to leave the system.

A model’s conversational fluency does not itself provide continuity after a session closes. Ira therefore separates several kinds of memory rather than treating the current conversation as the entire company record.

A salience gate decides what deserves retention. Without selection, persistent memory would accumulate incidental material alongside useful facts. When retained facts conflict, human corrections take precedence.

That precedence rule can be expressed independently of model behavior. In this illustrative TypeScript example, a human correction changes the payment schedule associated with a quote. A later conflicting extraction cannot overwrite it:

type Fact = {
  key: string;
  value: string;
  origin: "extraction" | "human-correction";
};

function reconcile(current: Fact, incoming: Fact): Fact {
  if (current.key !== incoming.key) {
    throw new Error("Cannot reconcile different facts");
  }
  if (incoming.origin === "human-correction") return incoming;
  if (current.origin === "human-correction") return current;
  throw new Error("Conflicting extractions require review");
}

const extracted: Fact = {
  key: "quote-2019-014.payment-schedule",
  value: "50% deposit",
  origin: "extraction",
};

const correction: Fact = {
  key: extracted.key,
  value: "30% deposit",
  origin: "human-correction",
};

const corrected = reconcile(extracted, correction);
const retained = reconcile(corrected, extracted);
// retained.value === "30% deposit"

The quote identifier and deposit amounts are teaching values. The important design choice is explicit authority: correction precedence is a rule the memory system applies, rather than a request for the model to remember which answer it should prefer.

Retention also has a maintenance cycle. Each night, Ira replays the day’s activity, consolidates useful information, searches for contradictions, forgets stale material, and turns work into reusable skills. In the morning, Doshi receives a dream report describing what the system consolidated, discarded, and worked out.

Doshi characterizes this as getting smarter overnight. The concrete mechanism is a scheduled process for maintaining memory and extracting reusable knowledge from experience; it is not another round of model training.

Each agent also has a soul file. Instead of relying only on generic helpfulness and harmlessness instructions, the file draws on the Jain principles of a family business spanning three generations. Doshi describes five inherited ideas translated into engineering rules, beginning with the recognition that no single source contains the whole truth.

The resulting constraints are practical: cross-check before speaking, avoid absolute claims, cite the document and its date, stay within the assigned job, and report the truth even when it is unpleasant. Nobody works alone. Doshi describes these as production guardrails: the system’s behavior should reflect how the business wants to make decisions, not merely produce persuasive answers.

Doshi reports a zero model-training bill. His account of the difficult work centers on teaching the company to remember itself: extracting and organizing private knowledge, rather than training a model. He says an agency quoted “two hundred and thirty grand,” while the internal build cost was “around thirty”; he does not specify the unit for the latter figure.

Doshi reports running costs of a couple of thousand dollars per month. No itemized accounting or workload accompanies the figures, so they describe his reported experience rather than a budget another factory can directly adopt.

Machinecraft extracted the architecture into BrainOS, presented as a forkable, empty nervous system. The agents, memory, dream cycle, and soul file provide the structure. An adopting company supplies its own history and knowledge; it does not receive Machinecraft’s populated memory.

That leaves the hardest responsibility with the business itself. Doshi’s closing resource is forkmybrain.org, but the invitation is to build a company’s own brain, not purchase his. He returns to Machinecraft’s lack of data scientists as the feasibility example: the reusable architecture can help, while the company must still decide what is true, which corrections matter, and how its accumulated knowledge should guide work.