A Factory Built Its Own AI Brain Without Data

Rushabh Doshi runs a hundred-person thermoforming machinery factory in India. Three generations of family business, no data science team, no machine learning budget. And recently, he stood in front of a room of AI engineers and described the system he built to replace what used to live exclusively inside his own head.

The story is worth your time—not because it's a triumph of technology, but because the problem it solves is one that most companies over fifty people have quietly given up trying to solve.

The Knowledge That Walks Out the Door

"We weren't scared of the competitors," Doshi said. "We were scared of forgetting."

That line deserves a moment. Machinecraft makes machines that heat plastic sheets and press them into shapes—farm trays, bathtubs, car panels, medical casings, packaging. Same core machine, seven different industries, seven different kinds of buyers. The knowledge of which customer wants what, what you quoted them four years ago, why their machine needed that one unusual modification—none of that lived in a database. It lived in people. And people leave.

Anyone who has watched a twenty-year sales veteran walk out the door on a Friday afternoon, taking half the company's institutional memory with them, knows exactly what Doshi is describing. It's not a technology problem. It's a human organization problem that technology might finally be equipped to address.

What Doshi built is called Ira. Not a chatbot, not an AI assistant bolted onto a CRM. He describes it as a company brain—a system of roughly three dozen specialist agents, each with a defined job, coordinated by an orchestrating agent named Athena.

Here is what that actually means in plain language: instead of one AI trying to do everything and doing it poorly, Doshi built a committee. Prometheus handles sales. Plutus handles pricing. Hephaestus knows the machine specifications. Vera fact-checks claims before they go out. Memnon, which Doshi calls his favorite, remembers every correction a human has ever made and ensures that correction sticks. "One agent, one job," Doshi said. "It's a team, not a hero."

Organized Memory, Not a Smarter Model

The part that surprises most engineers, by Doshi's own account, is what's under the hood. He trained nothing. No custom model, no proprietary AI, no server room of GPUs.

What he built instead is a very well-organized filing system that AI models can search. Years of quotes, drawings, payment schedules, email threads—hundreds of gigabytes of Machinecraft's private history—were broken into pieces and stored in a way that lets an AI retrieve the relevant bits when asked a question. Think of it as the difference between having the right answer in your files somewhere and actually being able to find it in under a second. The AI doesn't know more than what Doshi fed it. It just can't forget what it was told, and it can always find the right file.

This is the key architectural insight, and it's genuinely underappreciated: the value isn't in the model, it's in the memory. Off-the-shelf AI models—from providers Doshi uses in combination, selecting each for the task it handles best, he said—are capable enough. The scarce resource is context. What does this specific company know, about these specific customers, built up over these specific decades?

The system also has what Doshi calls a nightly dream cycle. Every night, Ira replays the day's activity, reinforces useful information, flags contradictions, and discards stale data. Every morning, Doshi gets a summary of what Ira consolidated overnight. It sounds like a metaphor. It is also a real engineering choice—scheduled processes that run consolidation logic while humans sleep, so the system improves incrementally without anyone manually retraining it.

A Conscience Written in Ancient Philosophy

This is the part I find genuinely unusual, and I'll be direct: I did not expect it to be serious.

Every agent in Ira operates according to a document Doshi calls SOUL.md—a set of behavioral rules drawn from Jain philosophical principles his family has held for three generations. The rules translate into concrete engineering constraints: no single source should be treated as absolute; always cite the document and the date; do your own job and stay in your lane; report the truth even when the truth is inconvenient; no agent works alone.

"Ancient philosophy running as guardrails in production," Doshi said.

I've seen enough AI systems hallucinate confidently, answer questions outside their domain with cheerful incompetence, and optimize for user approval over accuracy to recognize that these are not decorative principles. They are responses to real failure modes. The instruction to never state something absolutely without sourcing it is a direct engineering countermeasure against the most common way AI systems go wrong. The insistence that corrections always override original data is how you prevent a system from drifting back toward its mistakes.

Whether you describe it as philosophy or software design, the underlying logic is the same: systems that have no sense of their own limits are dangerous. Doshi gave Ira a functional sense of its limits.

What Doshi Built Versus What He's Selling

Here is where I need to be honest about a tension in this story, because Doshi is simultaneously reporting results and pitching a product.

The build cost him roughly $30,000, Doshi said—a fraction of what an outside agency quoted him to do the same job. An agency quoted him several times that amount for the same scope, he said. These are his figures, unaudited, and I cannot independently verify them. What I can say is that the architecture he describes—using existing AI models, existing database tools, and careful system design—is plausible at that cost range for someone who built it himself, over time, with deep knowledge of his own business.

That last clause is the one that should catch your attention.

Doshi has now packaged this architecture as something he calls Brain OS and is offering it through a venture called Fork My Brain, with the pitch that a specialized team can go into a company for a week, map the business, load the right files, and leave behind a working version of what he built. One week.

I have been covering technology long enough to remember when SAP implementations were supposed to take three months. When CRM rollouts were supposed to be plug-and-play. When digital transformation consultants promised they could modernize a company's operations in a single engagement. The one-week promise is almost certainly fantasy for any business of real complexity—not because the architecture doesn't work, but because the hard part was never the architecture. The hard part was Doshi knowing, with three generations of context, exactly which files mattered, exactly which relationships needed capturing, exactly what his system needed to know.

That knowledge is not transferable in a week. It's probably not fully transferable at all.

What the Factory Actually Proves

What Doshi built at Machinecraft is a real system solving a real problem. The daily operations it handles—outbound sales emails, pre-call account briefs, lead qualification, quote generation—are running, by his account, without a dedicated human sales team behind them. The golden rule he describes ("Ira drafts, human sends") is the right instinct: keep a human in the loop on anything that goes outside the building.

The honest version of the lesson here is narrower than the commercial pitch suggests, and more valuable for it. Doshi, as the third-generation owner of a business he knows better than anyone alive, built a system that captured what he knew and made it accessible and searchable and persistent. That is a genuinely useful thing. The question every other company needs to answer before getting excited about Brain OS is whether they have a Rushabh Doshi—someone who owns the knowledge deeply enough to know what to feed the system, what to leave out, and when it's getting things wrong.

If you do, this architecture is worth studying carefully.

If you don't, what you're more likely to build is a very expensive, very confident system that doesn't actually know your business—which is a problem the industry has been selling solutions to for thirty years.

By Bob Reynolds, Senior Technology Correspondent, BuzzRAG