Spend Less per Request
Run the workflow on Hookshot™ Models instead of sending every request to a large general model.
Get a frontier-level model post-trained on your cases, tools, and checks, at a fraction of the price. Call it through an API or use it in other Protege AI products.
Example figures. Accuracy varies by task and is measured through precision and recall.
Latency is how long it takes to answer.
Why Post-train for Repeated Work
Hookshot™ Models post-trains a model on your cases, decisions, and checks. Repeated work then costs less, comes back faster, and gives results your team can verify.
Run the workflow on Hookshot™ Models instead of sending every request to a large general model.
A private deployment returns more tokens per second than a general model on a shared API.
Tokens per second
Speed varies by model and deployment strategy.
Test each version on real examples with the checks your team already uses.
Hookshot™ Models
Test every candidate on the same cases, then host the frontier-level model that clears your bar.
Collect real examples of the work, the tools it uses, and the calls your team made.
Write checks for the result, the steps, and the answer format. Each check sends a reward back to the model.
Score every candidate on the same cases for accuracy, latency, and cost, then confirm the winner with a significance test.
Once it passes your checks on real test cases, host it on dedicated GPUs or Elastic Inference.
No ML Background Required
Anyone who knows the workflow can set up the cases and checks. Engineers can take it further in code.
Add real examples, then describe in plain English what a good answer is, how to reach it, and how to format it.
Switch to TypeScript when a workflow needs finer control. You don't need it to get started.
Call the finished model through an API or use it in Hookshot™ Agents. You don't have to run your own servers.
Hookshot™ Data brings operational context to training. Hookshot™ Agents take document review off your plate.
Model Services
We sit down with your team, learn the job, build the model, and help launch it. Then we hand the source code back to you.