Most "AI projects" start with a model and go looking for a use. We start from the task, the data and the number that has to move, and build the smallest system that reliably moves it. Then we hand it over.
A routing problem, an extraction problem and a Q&A problem are different problems. Each gets its own compact model, its own data and its own metric, instead of one giant model asked to do everything approximately.
Before any training starts, we agree on a gold set and a target. The engagement is judged on that number, in writing. No demos of cherry-picked examples.
Weights, data, pipeline, container. No license to us, no API key that can be revoked, no vendor that can deprecate your model. If we disappear tomorrow, it keeps working.
Two engineers sit with your domain expert and define the task precisely, "classify each incoming email into one of 14 queues", "extract these 12 fields from a supplier invoice", and what "correct" means. Together we assemble a gold set of a few hundred checked examples. That is the contract.
Good data is most of the work. We collect real inputs from your systems (or generate realistic synthetic ones when real data can't be shared), clean and de-duplicate them, and produce thousands of labeled examples, combining expert rules, the best available AI tooling and human review. Where policy demands it, every step runs offline inside your environment (see Security).
We train a compact open-weights model (typically 0.5-8B parameters) for your task alone, iterate against the gold set, and constrain the output to your schema, strict JSON, fixed label set, so downstream code never sees free-form text. On narrow tasks a purpose-built small model regularly matches or beats general frontier models; the eval report shows you exactly where it stands.
Delivered as a Docker/OCI image with an OpenAI-compatible HTTP API, so existing integrations work unchanged. Runs on an ordinary server, a regular CPU machine handles most workloads; for very high volumes a GPU card speeds it up. Nothing in the image phones home. We install it with your team and hand over weights, dataset, pipeline and eval harness.
Real traffic always surfaces cases the gold set missed. About a month after go-live we collect the errors and edge cases production produced, retrain, re-verify and redeploy. This round is part of the flat fee. After that, an optional retainer covers drift monitoring and periodic retraining, or your team runs the pipeline themselves.
Everything is transferable. No lock-in is the product.
Open-license base + your task-specific training. Stored where you say. Re-deployable without us.
OpenAI-compatible /v1/chat/completions endpoint. Drop-in for code that already talks to an API.
Re-run the numbers any time. Compare future models. Prove compliance.
The dataset, the config, the seeds. Reproducible by your own team.
Typical footprints, final sizing is part of the engagement.
| Task complexity | Runs on | Good for |
|---|---|---|
| Small | Any server, CPU only | Classification, routing, PII detection, short extraction |
| Medium | Ordinary server; a mid-range GPU card for high volumes | Structured extraction, entity linking, short-form Q&A |
| Large | Server with one GPU card | Long documents, SQL generation, multi-step extraction at high throughput |
We size the hardware during the engagement; most first tasks run on a server the client already has.
Same process, your data, your success metric, one fixed price.