Decision models, trained on your own decisions.
Our decision models read each ticket, document or claim once and answer in milliseconds, with a calibrated probability for every option. What they aren’t sure of goes to a person.
For support and ops triage, document review, and risk and compliance.
Samples, synthetic. Model time for three questions, measured: 18.4 ms at the median on one H100.
Ask in the shape the work already has.
A choice from your options, a score on your scale, or yes or no. Every answer comes back from the same pass with a probability for each option. There is no free text to parse, and nothing to hallucinate.
TKT-48213 Ticket
“Hi, I was charged twice for my October invoice and the second charge is still pending. Can someone reverse it before it posts?”
- teamchoiceWhich team should handle this?
- Billing0.97
- Account0.02
- Technical0.01
- urgentyes or noDoes it need an answer today?
- yes0.81
- no0.19
- frustrationscore, 1 to 5How frustrated is the customer?
- 40.58
- 30.27
- 50.11
- 20.04
Below the gate, a person decides.
Every model has a confidence gate, fitted on held-out data. When the best answer falls under it, our decision models don’t guess. The item goes to your team with its odds attached, and you set how much they decide on their own.
- Support triage
- With the gate set to decide 90% of tickets, 3.9% of those were wrong.
- Document review
- With the gate set to decide 90% of scans, none were wrong.
Measured on data it never saw.
Every model is scored on a test set held back before training, and the report ships with the model. These are ours, from our own runs.
| Model | Question | Before training | Before and after training, on a scale from 0 to 1 | Accuracy | Calibration error |
|---|---|---|---|---|---|
| Support triagemdl_712bfd99 | Team, 6 options | 0.614 | 0.892 | 0.046 | |
| Support triage | Urgent, yes or no | 0.862 | 0.970 | 0.031 | |
| Document reviewmdl_c07670ac | Document type, 4 options | 0.267 | 0.927 | 0.013 | |
| Document review | Paid stamp, yes or no | 0.913 | 1.000 | 0.042 | |
| Document review | Signed by hand, yes or no | 0.507 | 0.960 | 0.003 | |
| Intent routingMASSIVE, 3 runs | Intent, 20 options | — | 0.939 | 0.010 |
Support triagemdl_712bfd99
- Team, 6 options
- 0.892 accuracy
- Before training 0.614
- Calibration error 0.046
- Urgent, yes or no
- 0.970 accuracy
- Before training 0.862
- Calibration error 0.031
Document reviewmdl_c07670ac
- Document type, 4 options
- 0.927 accuracy
- Before training 0.267
- Calibration error 0.013
- Paid stamp, yes or no
- 1.000 accuracy
- Before training 0.913
- Calibration error 0.042
- Signed by hand, yes or no
- 0.960 accuracy
- Before training 0.507
- Calibration error 0.003
Intent routingMASSIVE, 3 runs
- Intent, 20 options
- 0.939 accuracy
- Before training —
- Calibration error 0.010
- Support triage
- 4,000 public banking support messages, each sent to one of six teams. Trained in 18 minutes on one L4 GPU.
- Document review
- 1,500 synthetic document scans, read as images. Trained in 5.3 minutes on one H100.
- Before training
- The same base model with no fine-tuning; for document review, reading the text alone.
- Calibration error
- The expected gap between stated and actual confidence (ECE). Intent routing: the public MASSIVE benchmark, the mean of three runs.
Three models. One API.
All three answer the same typed questions through the same API, in your cloud, on your own hardware, or in ours. Start with the strongest out of the box, and fine-tune a smaller one when you have the labels.
Fast
The cheapest decision, fine-tuned for each team’s work.
18.4 ms- Per million requests
- about $3.9
- Fine-tuning
- Per team
Edge
A larger model, fine-tuned for each team and sized for your own hardware.
41.6 ms- Per million requests
- about $25
- Fine-tuning
- Per team
Smart
The strongest out of the box, for work you haven’t labelled yet.
93.8 ms- Per million requests
- about $32
- Fine-tuning
- Optional
Measured in our benchmarks. Model time: one request of three questions, the median on one H100 in bf16, drawn on a 0 to 100 ms scale. Cost: that GPU at list price and full use.
Your decisions stay yours.
Trained only for you.
Examples you upload train only your models. They never go into a model anyone else uses.
Run it where you run things.
Serve it from your cloud, from your own hardware, or from ours, through the same API.
Keep the model and its evidence.
Export a model with its eval report: accuracy, calibration, the abstention curve, robustness and latency.
Text, documents and images.
Tickets, emails, scans, photos, screenshots and PDFs, read page by page.
Request early access.
We’re opening our decision models to teams in stages. Tell us where to reach you and what you’d like them to decide first.