Descriptions:
At an AI Engineer conference, Rushabh Doshi — founder of Machinecraft, a 100-person thermoforming machine manufacturer in India — describes how his company built a 36-agent AI system called Eira to encode and operationalize decades of institutional knowledge, entirely without a data science team, ML budget, or any agent framework.
The system ingests hundreds of gigabytes of historical quotes, engineering drawings, payment schedules, and email threads. Rather than fine-tuning models, the team chunks historical data into embedding vectors and a relationship graph, letting off-the-shelf models retrieve and reason over it at query time. Eira is organized as a cast of named specialist agents — Athena (orchestration), Prometheus (sales), Plutus (pricing), Hephaestus (machine specs), Vera (fact-checking), and Memnon (correction persistence) — that hold internal deliberations and surface a single coordinated output. All 213 tools are exposed over one protocol, and the non-negotiable rule is that Eira drafts while a human sends.
The architecture’s most distinctive feature is its layered memory system: working memory for recent context, pinned facts per contact, episodic conversation histories, relationship warmth scores that grow from stranger to trusted, and a nightly “dream cycle” that replays the day, resolves contradictions between stored facts, promotes useful patterns into reusable skills, and discards stale information. The talk is a compelling real-world demonstration of multi-agent systems delivering measurable go-to-market value in a traditional manufacturing context, with nine concrete daily jobs handled autonomously.
📺 Source: AI Engineer · Published July 11, 2026
🏷️ Format: Workflow Case Study
