Machinecraft built Ira as a company brain that preserves knowledge from documents, emails, quotes, relationships, and business processes.
2
Ira uses bounded specialist agents, layered memory, human approval, and verification rules instead of one general-purpose prompt.
3
Rushabh Doshi argues that companies should build forkable digital brains from their own information rather than buy generic AI assistants.
Summary
Rushabh Doshi describes how Machinecraft, a 100-person Indian thermoforming machinery company, built Ira to preserve knowledge that had previously lived in the heads of three generations of family members. The system ingests years of quotes, drawings, payment schedules, timelines, and email threads, then stores extracted meaning as vectors and relationships as a graph. Ira uses specialist agents for sales, pricing, machine specifications, fact checking, and corrections, with Athena coordinating their work. Its memory includes recent context, pinned facts, episodes, relationships, and a salience gate. A nightly dream cycle consolidates useful information, finds contradictions, forgets stale material, and produces a report. Doshi also describes a SOUL.md based on Jain business principles, including cross-checking sources, avoiding absolute claims, and reporting uncomfortable facts. Machinecraft built the system for around $30,000 after receiving an agency quote for $230,000, then turned the architecture into the forkable Brain OS project.
Machinecraft needed to preserve knowledge that lived in a few people's heads
Rushabh Doshi says Machinecraft's real value is its accumulated knowledge: customer histories, old quotations, unusual machine changes, and the reasons behind past decisions. For three generations, that knowledge lived mainly in his grandfather's, father's, and his own heads. Employee departures repeatedly removed part of the company's memory. The company was less worried about competitors than about forgetting how it worked. Doshi decided that ordinary documentation would not be enough because people often do not read static documents. He wanted a company twin that could hold and use the knowledge directly.
The company brain is built from organized memory rather than model training
Machinecraft gave Ira years of quotes, drawings, payment schedules, timelines, email threads, and hundreds of gigabytes of private company history. The system did not train a new model or use fine-tuning. It broke the material into chunks, used off-the-shelf models to extract facts, stored meanings as vectors, and recorded connections as a graph. Doshi says the system's advantage is not a smarter model. It is well-organized memory grounded in the company's own history. That grounding also lets it distinguish among very different customer contexts, from hydroponic farm trays and spa bathtubs to EV panels and medical casings.
Ira is organized like a biological system so it can stay coherent
The team stopped treating Ira as ordinary software and gave it a structure based on biology. It has senses to identify who it is speaking with, a gut that digests documents into facts, memory, a dream cycle, and an immune system for bad information. Doshi's reason for the metaphor is practical: evolution has spent a long time dealing with continuity and coherence. The system's components divide those responsibilities instead of leaving every task to a single prompt. This design frames ingestion, memory, correction, and verification as connected parts of one operating system.
Specialist agents work as a coordinated team instead of one general prompt
Ira uses bounded specialists with one job each. Athena coordinates the work, Prometheus handles sales, Plutus handles pricing, Hephaestus knows machine specifications, Vera checks facts, and Memnon protects corrections. Doshi says a prompt that tries to do everything usually performs each task badly. Athena brings the relevant specialists into a meeting, where they can disagree before producing one answer. The result is a system that treats business work as collaboration among agents. The description identifies 39 bounded agents, while the talk gives examples of the specialist roles.
Ira covers daily go-to-market work while keeping a human in control
The system handles outbound emails, account briefs, quotations, outreach triage, dead-lead revival, inbound replies, and early checks on whether a company is a fit. Doshi operates it through one Cursor tab. Ira can search the knowledge base, read the inbox, draft an email, build code, and present the result for review. The system connects vector, graph, CRM, and other databases with multiple model providers and communication tools. Its capabilities are exposed as 213 tools through one protocol. The operating rule is simple: Ira drafts, and a human sends.
Layered memory and nightly consolidation give Ira continuity
Doshi describes several memory layers: working memory for recent context, pinned facts about a person, episodes that preserve conversations as stories, and relationships that can move from stranger to trusted. A salience gate decides what deserves storage. When facts conflict, corrections take priority. Each night, Ira runs a sleep cycle that replays the day, keeps useful information, searches for contradictions, removes stale material, and turns work into reusable skills. Doshi reads a morning dream report that describes what was consolidated, discarded, and learned. The design aims to preserve continuity without inventing information.
The system's SOUL.md turns Jain principles into operating rules
Ira's agents have a SOUL.md based on principles from Doshi's Jain family business. Those principles become concrete rules for production behavior. No single source has the whole truth, so the system cross-checks before speaking. It avoids absolute claims and cites the document and date. Each agent does its own job rather than taking over another agent's responsibility. It reports the truth even when the result is uncomfortable. Agents also work together instead of operating alone. The philosophical document is therefore used as a set of guardrails for retrieval, attribution, specialization, and reporting.
Brain OS packages the architecture as an empty system companies must fill themselves
Doshi says an agency quoted $230,000 to build the system, while Machinecraft built it for around $30,000, with ongoing costs of a couple of thousand dollars per month. He says compute was not the expensive part. The harder work was teaching the company to remember itself. Machinecraft extracted the architecture into Brain OS, which includes agents, memory, a dream cycle, and a soul file without any company-specific content. Each organization must supply its own documents, relationships, processes, and truth. The proposed model is to build a company brain from the inside out rather than buy a generic dashboard.
"The brain isn't a smarter model. It's actually a really, really well-organized memory."03:10
Who should watch
You run a company where customer history, quoting decisions, and operational knowledge are spread across people, inboxes, and old files.
Your team is deciding whether to build a multi-agent system around company data instead of adopting a generic assistant.
You want a concrete account of human approval, memory, corrections, and organizational rules in an AI system used for business work.
