How we partner
Six weeks from archive to asset
Every engagement runs the same fixed-scope sprint. You know the price, the timeline, and the yardstick before anything is trained.
WEEK 1–2
Evaluate
We scope one high-value workflow with your domain experts and build its eval harness from your archive. The best frontier models are scored against it — this baseline is the go/no-go gate for everything that follows.
Deliverable
Your eval harness + a written frontier-baseline report.
WEEK 3–4
Train
Reinforcement learning and preference optimization on open weights, inside your infrastructure or an agreed sovereign region. Daily eval scores, full rollout traces, no black boxes.
Deliverable
A specialized model beating the baseline on your eval.
WEEK 5
Deploy
The winning checkpoint promotes to production in your VPC, datacenter, or national cloud — in the same harness it trained in, sized to your latency and throughput needs.
Deliverable
A production endpoint inside your boundary.
WEEK 6
Hand over
Weights, harness, eval suite, and runbooks transfer to you as contracted IP. We train your team to operate the stack, and agree (or don't) on a continuing improvement cadence.
Deliverable
You own everything. Ongoing partnership is optional.
Engagement principles
Eval before invoice.
The first deliverable is a measurement, not a model. If frontier APIs already win on your workload, the engagement ends there — and you keep the eval.
You own the weights.
Every contract transfers the trained model to you as IP. No per-token tax on your own intelligence, no dependency on our uptime.
We keep nothing.
No copies of your data, no rights to your model, no training our next customer on your alpha. That is the entire point of the company.
Fixed scope, fixed price.
Six weeks, one workflow, one number agreed up front. Expansion happens because the first model earned it.
Is this you?
A good fit looks like this
- 01
You operate in insurance, logistics, or legal ops — or another regulated, document- and decision-heavy business.
- 02
Your data cannot leave your jurisdiction, or you're unwilling to hand your operating history to a US frontier lab.
- 03
You have an archive of decided cases — settled claims, completed routes, negotiated contracts — with known outcomes.
- 04
One workflow is expensive enough that a few points of accuracy move a real KPI.