BCUB3 Lab
This page exists for two reasons. First: explain in plain language, without formulas, what lives under the hood of the models we deploy with our industrial clients. Second: present our proprietary patented algorithms — for model efficiency and for industrial process control — that make AI more efficient and more deployable on a shop floor than in a data center.
Agentic kernel orchestrating swarms of pods (MCP-compatible agentic CLIs), with three-tier memory and prompt evolution via GEPA — without mutating the kernel itself.
Read technical documentation → Models · Infra · Stack5 proprietary models compared (Claude, GPT, Gemini, Mistral, DeepSeek) — context windows, strengths, weaknesses. Plus the GPU, edge and open-source inference stack we deploy on the shop floor.
See the stack → Patented algorithmsProprietary patented algorithms: AI model efficiency and industrial process control. Explained without formulas.
Discover our patents → Founding doctrineWhy our systems get better when the factory shifts. An antifragile thermostat in Nassim Taleb's sense, transposed to industrial process control.
Understand antifragility →Section 2
All our tutorials run 100% in your browser. No data leaves your machine. Drop a CSV, tweak the parameters, see what happens.
Section 3
We file patents because our industrial clients need to build their AI strategy on foundations with clear intellectual property. Two families: AI model efficiency and industrial process control. Here are the big ideas, explained without formulas, using everyday analogies.
What it is, in one sentence
A proprietary building block that lets small models learn more efficiently, getting closer to the performance of far larger ones.
Why it matters in industry
Compact models deployable on the edge (cameras, sensors, robots), where memory and compute budgets are tight — without sacrificing quality.
Analogy
Like a better-tuned engine: for the same displacement, it gets more out of every litre of fuel.
What it is, in one sentence
A controller that tells plain measurement noise apart from a real process drift, and only acts when it is statistically justified.
Why it matters in industry
Less scrap and less variability, without an expert operator constantly arbitrating — the system re-centers itself when the environment shifts.
Analogy
Like a smart thermostat: it ignores a brief draft but corrects firmly on a true trend.
Section 4
On an industrial site, the environment is never static: raw materials vary, sensors drift, operators rotate, seasons shift the temperatures. A rigid system degrades. A robust system resists. An antifragile system — in Nassim Taleb's sense — learns from stress and gets better. That is the principle behind our R&D.
01
A robust thermostat holds steady despite fluctuations. An antifragile thermostat learns from those fluctuations to better anticipate the next one. Our proprietary controllers go beyond robustness: they use every drift as training signal, without human intervention.
02
Rather than freezing an AI strategy, we let it mutate under measured constraints: user satisfaction, latency, token cost, error rate. The approach draws on recent academic work in reflective prompt evolution (GEPA, 2025) and genetic programming. The best variant survives; the others are retired.
03
We separate two layers in every system we ship. The kernel — safety invariants, output contracts, business policy — is engraved and never mutates. The harness — phrasings, hyperparameters, heuristics — evolves continuously. This split guarantees operational stability while still allowing learning.
Multi-pod harness in continuous evolution
The agentic runtime that orchestrates swarms of agent pods, persists knowledge, and pilots online probabilistic controllers. Technical documentation, in English and French.
A question, a concrete case?
If one of these topics sheds light on a problem you're hitting in the field — defect detection, machine drift, non-standardized documents — let's talk.
Discuss a concrete case