Custom models trained for critical decisions.
We train our own models, each on the documents one line of work runs on, then run them for you. You pay for inference.
One decision, made well.
A model learns one call your team makes again and again, from the thousands of times they already made it. Then it makes that call in seconds, and hands the unclear ones to a person.
Our models.
One is ready to use, three are in training and four are planned. Each reads one line of work.
Vulcan 2.5
Turns lab reports, spec sheets and vendor packets into a product record, every value tied to its page.
Altadena 1.5
Fills the benefits breakdown, about 290 boxes, and leaves blank whatever the payer didn’t say.
Pasadena 1.0
Settles whether each patient on the schedule is covered, straight from the payer portals.
Blob 2.0
Reads hours from time sheets, timecards and emails into a pay line for each worker.
Ridge 1.0
Hands the asset manager every date in the rent rolls and leases, with the page it came from.
Arroyo 1.0
Builds the advisor a report with every figure linked to its custodian statement.
Phantom 1.0
Drafts from a case’s intake notes and orders. Nothing is advice until an attorney signs it.
Svengali 1.0
Reads an event’s contracts and expenses into holds to confirm and settlements to check.
Better at less.
A general model answers questions on everything from lab reports to poems. Ours learn one line of work.
93.0%
Altadena 1.5
Of the boxes right, on ten breakdown sheets it had never seen. Still in training.
70.9%
OpenAI’s GPT-6 Luna
On the same ten sheets. We published the test.
Why we train our own.
- 01
More accurate.
Trained on the documents one line of work runs on, not on everything at once. 93.0% to GPT-6 Luna’s 70.9% on our breakdown test.
- 02
Cheaper.
A small model built for one job costs a fraction of a frontier model to run, and you pay only for what it answers.
- 03
Sovereign.
Open-source weights, hosted by us, not a frontier lab. Your records train your model and no one else’s.
- 04
Same answer, every run.
Same document, same version, same answer. Every write records the model behind it, so any reading can be traced and repeated.
How we build it.
Sign
We agree on the decision, and sign an MSA, a DPA and, for health data, a BAA.
Train
We fine-tune an open-source model on your past decisions and the records behind them.
Measure
We test it on cases it has never seen, next to your team, before it makes a single call.
Run
We host it, watch it and retrain it as your business changes. You pay for what it answers.