Turn how your team works into an AI model you own.
Teach LUK a task once. It captures the clicks, corrections and explanations, and trains an open-weight model that runs on your own hardware.
Backed & trusted by
Teach by doingShow it once, from the first click to the last correction.
No SOPs, no workshops. One person does the task as usual while LUK runs in the apps chosen for it, and every click, file, correction and spoken explanation becomes training data.
What you ownTurn one teaching session into a model that is yours.
Not a configuration on someone else's model. The weights, the training data and the evals live on your hardware, and carry over to better open models as they ship.
| Weights | Size | Runs on | Eval |
|---|---|---|---|
| invoice-entry.safetensors | 4B | srv-ops-01 | 90% |
| invoice-entry.base | 4B | srv-ops-01 | 70% |
Your scoreboardKnow how good it is before it touches real work.
A 4B model trained on our pipeline landed within three points of Claude Opus 5.5 on 98 real French electricity bills it had never seen, at about 45× lower cost per document.
Field accuracy, zero invented values. One document task, not a general benchmark.
On your hardwareRuns on your servers. No GPUs needed to teach.
Teaching runs on a normal Windows laptop. The trained model is small enough for your own server. We host only if you ask us to, at cost plus 15%.
Teach
Someone does the task as usual while LUK runs in the apps chosen for it. They can talk through the tricky parts.
luk teach invoice-entryTranslate
Clicks, files, corrections and explanations become training data and evals for that task.
examples + evals, yoursOwn
A small open-weight model is trained on it and runs on your hardware. The weights are yours.
invoice-entry.safetensors
Meet LUKYour new hire that learns your workflows and automates them.
LUK captures while someone is teaching, in the apps chosen for the task.
Learn more
What LUK captures
- Clicks and UI actions in the apps chosen for the task
- What is typed into fields, and the files used
- The spoken explanation, if the person talks it through
- Corrections, which become evals
What it never captures
- Apps outside the task, or the whole desktop
- Webcam or background audio
- Anything while teaching is paused or off
- Individual scores, rankings or emotion inference
Built for teams that answer to a works council and a DPO.
Hosting is yours or ours, your choice. We are happy to walk your works council or DPO through exactly what LUK does and does not record.
Ask about a works council walkthroughQuestionsWhat ops and IT ask first.
Data, works councils, GPUs, cost. For anything else, write to info@cledion.com.
Does our data leave the building?
No. LUK captures on the device, redacts and blurs sensitive content before anything is stored, and keeps the data encrypted. Training data, evals and the trained model live on your own infrastructure. We host only if you ask us to.
What does LUK capture, and what does it never capture?
Only while someone is teaching a task, and only in the apps chosen for it: clicks and UI actions, what is typed into fields, the files used, and the spoken explanation if they talk through it. Never apps outside the task, the whole desktop, the webcam or background audio.
Will our works council or DPO accept it?
LUK is built for that conversation. Teaching is opt-in and can be paused at any time, it runs only in the apps for the task, and it judges outcomes, never people: no individual scores, no ranking, no emotion inference. We are happy to walk your works council or DPO through it.
Do we need GPUs?
Not to teach. LUK runs on a normal Windows laptop with about 5 to 6 GB of RAM. The trained model is small and runs on your own server. If you have no GPUs, we can host it for you at cost plus 15%.
What does it cost?
No per-token bills and no per-agent fees. Capture is priced per seat, one seat per person who teaches LUK. Training and running a model is priced per process. We are onboarding a few teams at a time from the waitlist.
Get early access.
We're onboarding a few teams at a time. Join the waitlist, invite a colleague, and move up.