AI consultancy for companies building real products — where software, firmware, and hardware meet intelligence.
The numbers below come from a production multi-device AI system running in our own lab — local language models, device coordination, and voice control across a mesh of consumer-grade hardware. It is the kind of system we build for clients, measured rather than described.
| llama3.2:3b | 36.7 tok/s |
| qwen3:4b | 30.1 tok/s |
| deepseek-r1:8b | 17 tok/s |
| qwen3:8b | 16.6 tok/s |
LAST MEASURED 15 JULY 2026
You build a physical product and the board wants AI in it. We turn that into a costed, buildable roadmap — including the honest answer when the right amount of AI is less than expected.
grounded in shipping firmware, not slideware
Language models, classification, and decision-making running on the device itself — no cloud round-trip, no per-query bill, no customer data leaving the product.
see SHT 2 — those are our own measurements
The unglamorous part most AI consultancies can't do: making intelligence coexist with real-time constraints, watchdogs, OTA updates, and the rest of an embedded stack.
electronics, firmware & software design background
Condition monitoring, anomaly detection, and process optimisation from the sensor data your line already produces.
R&D experience across consumer & industrial products
Inspection, counting, presence, and safety applications — scoped to run on hardware that matches the budget of the product it ships in.
edge-first: the camera and the model share a box
Which model, at which size, on which hardware — answered with benchmarks rather than vendor claims. The right model is often smaller and cheaper than the default choice.
we rank models by measured task performance
Where the experience comes from
electronics, firmware & software design roles — preceded by 20 years in contract R&D
Write it the way you'd explain it to a colleague — our assistant structures it into a scoped brief on the spot, before you decide whether to send it.
BRIEF — AWAITING INPUT