Early access

Distill expensive
traces. Own the
model.

Converra takes one workload you run on a frontier model and tests whether a small open model can do the same job. You get the evidence either way.

Read the 48-part series

One workload. One answer.

  1. 01

    Picks one workload

    The expensive, repetitive one. One task, and the quality bar your team already holds it to.

  2. 02

    Builds the test from your traces

    Saved runs become the question set. Your team confirms which answers count as correct. A protected set is held back before any training.

  3. 03

    Trains and compares

    A small open-weight model trains on the reviewed runs, then answers the protected set next to the model you run today.

  4. 04

    Returns a verdict

    Pass, fail, or not enough evidence. A pass earns a trial on live traffic, and you approve the switch.

What you get.

What you own
The model, the adapter and the training dataset. They are yours under our Terms and are never shared with another customer.
What trains it
Your own saved runs, reviewed by your team. Insights point out what went wrong; they never stand in for a correct answer.
How you know
Both models answer the same protected questions and are scored the same way. You see every answer, not only the total.
Where it runs
Your provider account or ours, decided with you before anything is trained. Providers are barred by contract from training on your data.
Who does the work
Oren Cohen, Converra's founder, runs every engagement: scoping, data review, training and the verdict.

What we don’t promise.

  • That the small model passes. Some workloads need the frontier model. When that is the answer, the test says so.
  • A savings number before the test. Serving cost depends on your volume. At low volume, a dedicated small model can cost more than the API you use now.
  • A production result from a lab score. A passing test earns a trial on live traffic. Real users decide the rest.

From production · Verified on real traffic

22 → 0

Invented pricing, contact emails and product features, across 97 sampled conversations after the fix.

That result came from a prompt fix, not a model transfer. It is here because the same production verdict judges a model switch.

The method, in public

48 parts on moving a production agent from a closed model to an open one.

Data selection, leakage, judging, serving cost and rollback. Written from the runs that failed as well as the ones that held.

Read the series →

Start with one workload.

Thirty minutes to pick the workload and the bar. You see the test plan before anything is trained.