Fine-tuning
A model adapted to your task with curated examples from your own work.
Enterprise AI / Specialized models
Models trained for your job.
When a general model isn’t accurate enough for your documents, your field or your language, we adapt one to the task and prove it on your own cases.
What we build
Where we usually start.
Retrieval done properly
Often the right answer is better retrieval rather than training. We test both and keep what scores higher.
Evaluation
A fixed set of your real cases, scored blind, with repeat runs to check it gives the same answer twice.
Non-English languages
Speech and text models for languages general AI handles badly.
How we measure it
A baseline before, the same numbers after.
Every model is scored on your own cases before it ships, and every later change is scored on the same set.
- 01Accuracy on your test set
- 02Consistency across repeat runs
- 03Latency
- 04Cost per request
How it runs
Four steps, the same on every project.
- 01
Map the work
We sit with the people who do the work today and map the process as it actually runs.
- 02
Measure the baseline
Time, errors and cost for the task as it is now, written down before we build anything.
- 03
Build in small releases
Working software early, reviewed with you as it develops.
- 04
Report the same numbers
After launch we measure the same things again, so the result is a number, not an opinion.
From the blog
Guides on this.
Security and compliance · 7 Aug 2026 · 6 min read
My Supabase RLS Policy Is Not Working
A debugging order for Supabase row-level security when policies seem ignored or block everything, including the service-role trap and the tests that prove it.
Security and compliance · 6 Aug 2026 · 6 min read
Someone Found My API Key. What Do I Do Right Now?
A calm, ordered playbook for the hour after you discover an API key leaked: rotate the right things first, read the logs, and purge the key from history.
Security and compliance · 4 Aug 2026 · 8 min read
My App Works for Me but Not for Other Users
A diagnostic guide for the app that works perfectly on your account and breaks for every other user, including the security problems hiding behind the symptom.
Questions
What buyers usually ask.
Often not. Many jobs work well with a hosted model and good retrieval. We recommend training only when the baseline shows a general model can’t meet the bar.