← back
The Factory That Dreams: 39 AI Agents, No Framework - Rushabh Doshi, Machinecraft
Takeaway
A useful company agent system can emerge from organized institutional memory, narrow specialists, and persistent human corrections.
Summary
- Machinecraft built Ira to preserve decades of quotations, drawings, emails, and customer knowledge using off-the-shelf models, vector search, and a relationship graph.
- Specialist agents handle coordination, sales, pricing, machine specifications, fact-checking, and human corrections, with 213 tools exposed through one protocol.
- Layered memory stores working context, pinned facts, episodes, and relationships; nightly consolidation detects contradictions, removes stale information, and extracts reusable skills.
- The system supports outreach, account briefs, quotations, and inbound work under a human-sends rule; its reusable architecture is offered as Brain OS.
multi-agentagent-memorymanufacturing
Original description
Most AI demos are built around a toy workflow. Ira was built around a factory. This talk is the story of how a third-generation Indian machinery company built a multi-agent operating system that helps run sales, business development, recruitment, quoting, marketing, production context, email workflows, and organizational memory. Ira is not a chatbot and not a wrapper around a single framework. It is a company brain: 39 bounded specialist agents, Athena as orchestrator, a 17-stage request pipeline, Qdrant for document memory, Neo4j for relationships, Mem0 for long-term semantic memory, Postgres for CRM and recruiting data, Redis for coordination, and Cursor as the operating cockpit. The deeper lesson is architectural: companies do not need generic AI assistants. They need digital brains grounded in their own documents, relationships, processes, and values. I will show how Ira ingests company files through a "digestive system", routes work through a pantheon of agents, verifies claims through immune-system style guardrails, learns through memory and corrections, and "dreams" through a nightly consolidation cycle. I will also explain why we gave Ira a SOUL.md: a philosophical constitution based on Anekantavada, Syadvada, Svadharma, and operational truthfulness. The talk ends with the Fork My Brain thesis: the right way to build company AI is not to sell another SaaS dashboard. It is to send a special-ops AI team inside a company for a week, map the business from the inside out, ingest the right files into Qdrant and Neo4j, wire the operational databases, and leave behind a forkable digital brain that employees can run through Cursor and LLMs. Speakers: Rushabh Doshi (Machinecraft / Fork My Brain): Rushabh Doshi builds and operates Ira, a multi-agent AI operating system for Machinecraft, an Indian thermoforming machinery manufacturer, combining Cursor, LLMs, retrieval, memory, and business operations into one living company brain. LinkedIn: / rdd0101 GitHub: https://github.com/doshirush1901