Research
The Vital Matrix
We stack AI, trust, physical signal, simulation, and robotics when the problem actually needs it—usually not all at once, and rarely in the order the hype cycle suggests.
Lab log · mm/yyyy
Notes from the messy middle
Honest chronology: what failed in the field, what we stopped pretending about, and the habits we kept because they saved our clients money.
The LLM demo died on factory Wi‑Fi
We brought a laptop-only “ask the manual” prototype on site. Worked in the conference room; on the floor, roaming APs and packet loss made answers feel broken. That afternoon we sketched the first Vital Matrix note: intelligence has to sit next to signal and integrations, not pretend the network is clean.
MQTT looked easy until the broker didn’t
Edge spike with a cheap gateway and three sensors. Topics were fine; certificates and firmware update paths weren’t. We started a shared doc titled “integration is the product” — half the bugs were contracts and retries, not model accuracy.
A four-week MVP that was mostly APIs
Startup pilot: “computer vision on the line.” Shipped a workable detector in two weeks; the next two went to SAP timestamps, a flaky handheld scanner, and who was allowed to see which alert. We stopped being surprised when the ticket board said “IoT” but the work was Nexus-shaped.
Five layers on one whiteboard
Retrospective across six engagements: the same five buckets kept appearing — data/ML, physical signal, glue between systems, rehearsal/visualization, shipping to prod. We named Core, Sense, Nexus, Sim, Forge so PMs could argue scope without swapping vendor decks mid-meeting.
Handoff doc that actually compiled
After a Sim-heavy sprint, the client’s engineer cloned the repo, ran docker compose, and hit the twin viewer on the first try. We made that sequence non-negotiable: README, env template, and a 15-minute screen recording. “Works on our machine” became a failed gate, not a punchline.
Security review before the hero slide
Enterprise deal: legal asked for data flow and key custody before the exec demo. We moved IAM, retention, and audit logs into week one of the sprint instead of week four. Same deliverables, fewer panicked threads the night before sign-off.
Orchestration spikes, still two-week proofs
Running experiments with multi-agent task graphs and tighter Sim loops for capex decisions. Open questions: who owns failure when two agents disagree, and how much twin fidelity is enough to trust a number. Rule unchanged — pick a slice, instrument it, hand it off if the bet clears.