From the Lab
What we build, how we build it, and what we learn along the way.
How We Deliver Story to a Factory Agent
Retrieval is useful, but operational AI systems need more than a bag of context. In Factory Agent, we separate stable identity, tribal knowledge, bounded snapshots, conversational contracts, and living story objects so the model receives the right shape of truth for the job.
Three Runtime Cross-Fertilization: How Fast Loops, Live Conversations, and Overnight Reasoning Improve Each Other
Most AI systems pick one runtime and call it a product. We built a factory intelligence system where cheap loops, user-facing conversations, and deep offline reasoning continuously teach each other.
Why We Built a Factory MCP Instead of Letting Models Touch Raw Data
Large models are good at language, not at safely navigating factory data. We built a bounded factory MCP layer so agents ask better questions, use the right abstractions, and stop taking dangerous shortcuts.
You Can Port the Harness, Not the Relationship
AI coding systems are supposed to be portable. In practice, switching partners means leaving behind a co-adapted working relationship — the model, the harness, the interface, the rhythm, and the workforce you designed around your own mind.
The Studio Harness: Running a Real Engineering Studio With Persistent AI Memory and Verification
The hard part of an AI-native engineering studio is not generating output. It is deciding what becomes durable truth, what must be re-verified, and how work survives across sessions without turning into stale confidence.
Why Traditional Industry Needs AI Infrastructure, Not Just AI Features
Most industrial AI products lead with features: chat, dashboards, copilots, alerts. The real leverage comes from the infrastructure underneath them - machine connectivity, bounded data contracts, runtime discipline, and delivery into the actual operation.
From Tribal Knowledge to Operational Intelligence
Factory data becomes truly useful only when it is joined with the informal rules people carry in their heads. The path from tribal knowledge to operational intelligence is not capture alone - it is translation, testing, and promotion into the system.
A $350 Raspberry Pi Replaced a Failing Industrial Sensor
How classical computer vision on a Pi 5 outperformed an industrial photoelectric sensor for counting bags on a packing conveyor — and why we didn't use a neural network.
Reverse-Engineering a Proprietary Industrial Protocol on a Raspberry Pi
How we connected to optical sorters using a proprietary binary protocol — no manufacturer rep, no SDK, just docs and trial and error.