The core conversational AI runs locally. MIKE answers, understands, performs controlled actions and keeps verifiable evidence of critical actions.
Start with a 30-minute conversation. Response within two business days; scope and fees agreed in writing.
Pilot stage. The interface shown here is a dated capture, not a live customer deployment.
That last number is not a typo.
Replay a bounded example of an incoming call.
This illustration replays pre-computed call steps. It is not a live call or a receipt from a customer deployment.
STATUS AND LIMITS
Voice latency, Québec French over the phone and simultaneous calls still require validation. The dated measurements and their limitations are public.
Read the measured evidence ↗A pilot is a fixed scope with a fixed end date, agreed in writing before anyone starts.
Fixed fee, quoted in writing after a 30-minute scoping call.
If the measured absorption comes in under 30 %, we say so in writing and do not propose a deployment.
This is a small independent practice, not a platform with a sales quota.
Deployment is quoted against the scope the pilot actually measured.
A fixed fee, quoted in writing after a 30-minute call.
You get the figure before you commit to anything.
Shadow mode means MIKE proposes and a human sends.
The journal tells you exactly which action and who approved it.
MIKE's proposed distinction is a local runtime with evidence on your disk.
If your bottleneck is writing, you don't need MIKE.
MIKE runs on your hardware, in open formats — SQLite, YAML, NDJSON.
What you would lose is support and new skills, not the operation.
A single workstation with a consumer GPU handles a small business workload.
Nobody, as a paying customer.
Being among the first is a real risk, and the pilot is structured so that it is a small one.
It replies in whatever language the customer wrote in.
One real workflow, measured in shadow mode. Tell us what slows your team down; we will scope the pilot with you.
Send directly to the private F51 laboratory inbox. You will see a receipt once your request is saved. We use your contact details only to discuss this pilot.
MIKE answers the phone in French and English, understands what the caller wants, writes into MIKE’s own agenda and keeps a receipt of every action it takes.
Real-time voice architecture under validation, with continuous listening and speech interruption.
Move the three sliders to your real situation.
Loaded cost means salary plus payroll taxes, benefits and overhead.
You don't hire people to answer questions. You hire them to close loops.
It plans, acts, verifies, remembers and follows up.
Shaded = the only moment that needs a human.
A live MIKE instance answers questions about the product below.
What it answers: what MIKE does, how a mission runs, how a pilot is priced.
This demo sends your question through Cloudflare to F51 Labs' machine. Replies are generated and can be wrong. Do not enter private information.
Ordered steps, the tool each one needs, what gets checked, where you sign off.
Runs entirely in your browser.
Every row carries a real SHA-256 hash computed in your browser.
| Time | Tool | Action — editable | Result | Check | Hash (prev → this) |
|---|
Here is the honest split.
Designed with Québec's Law 25 in mind.
MIKE grew out of independent work on machine learning, agent systems and verification.
Brazilian, in Montréal. Self-taught.
At eight years old I took apart the family VCR and got it working again.
The love of machines never went away, but life put it to sleep for a long time.
At fifty-one I started studying machine learning.
Self-taught. Published research with a DOI. I build systems whose results are meant to be measured, reproduced and verified.
I am not here for the money.
This is the discipline that runs inside F51.
Written down, not measured.
Measured, but not yet checked against a control.
Measured, reproduced, checked against an active negative control.
Five lines, at five different stages.
Large MoE models activate only a fraction of their weights per token.
The headline figures were withdrawn: no file supports them.
A prefix-cache benchmark reported a 57 % latency reduction. The equivalence check killed it.
The real bottleneck was hundreds of CPU↔GPU synchronisation events.
An agent that runs bounded business missions and stops at human approval gates.
Several lightweight predictor voices propose likely routing trajectories ahead of the token.
Target: 235B–671B MoE models at usable speed on modest hardware.
An internal audit found three quarters of its experts had been permanently frozen.
The checkpoint was not patched. The lineage was restarted.
The business reading explains what happens.
Each one declares its own autonomy level and status.
None of this runs on rented A100s.
MIKE was designed to run on a single workstation a small business can own.
The only durable asset is whether the claims hold up.