A green test suite missed it
Churn serving parity
A plausible patch passed software checks while online and batch scores disagreed. The ML checker rejected it.
Historical genuine recording · unisolated test runnerInspect the rejected patch
Repository repair for ML engineers
The ML coding agent that checks beyond green tests.
Turn a failing ML repository into a reviewable patch, with independent evidence that the model still behaves correctly.
Inspect the code. Inspect the checks. You control the merge.
Genuine recording. No account or API key needed to watch.
churn-serviceRecorded evidence
candidate.patch
+ "contract_monthly", "auto_pay", "internet_fiber", "paperless", "usage_gb",Green tests. Wrong behavior.
Online and batch scores disagreed on 200 rows. This patch was rejected.
Historical unisolated test runnerRead the actual receipt
Built around your engineering workflow
Python repositoriesNemotron reasoningIndependent verificationReviewable Git diffs
Genuine hosted repairs use Nemotron through Nebius Token Factory, with isolated execution in Nebius Sandboxes. Historical recordings retain their original provider labels.
The gap after “tests passed”
A working pipeline can still make the wrong predictions.
Serving code can build features differently. A repair can learn from leaked labels. An improved average can conceal a weaker result for one group.
Autopilot investigates the code, tries bounded repairs and asks a separate verifier to check the evidence before presenting a finalist.
See the trust boundariesEvidence you can open
A claim is only as good as its receipt.
6sample repositoriesSix distinct ML failure stories
5/5wrong-green patches rejectedAuthored checker proofs, not an agent success rate
Opensource, diffs and receiptsGPL-3.0 · inspect the implementation
An early product with inspectable evidence. We do not claim customer adoption, production case studies or a completed comparative benchmark.
A tempting shortcut is refused
Refund label leakage
An authored wrong patch gains access to information unavailable at prediction time. The fixture proves the leakage gate rejects it.
Authored checker proof · not an autonomous repairInspect the leakage proof
One average can hide a regression
Credit group recall
A higher overall score does not excuse weaker recall for a protected group. Declared segment checks evaluate that trade-off.
Authored checker proof · declared groups and limitsInspect the group-check proof
From failure to a decision
The agent does the investigation.
You get the evidence.
Every stage produces something concrete to review.
Set the boundaries
Pin a repository snapshot and choose the failing test, permitted edits and ML requirements.
Investigate the cause
Nemotron compares hypotheses against repository evidence. Relevant official documentation can inform the next edit.
Try isolated repairs
Candidate branches produce actual source edits and regression tests under bounded tool and token budgets.
Check code and model
A separate verifier checks software behavior, declared ML invariants and the available evaluation data.
Review and decide
Read the diff and both receipts. Download the patch or explicitly open a draft PR. Nothing merges automatically.
Built for the review you actually do
Less guessing.
A clearer handoff.
Designed for the ML engineer fixing a serving incident, the maintainer reviewing a generated patch and the reviewer asking what a green check really proves.
Fix your repoA diff, not just an answer
See the exact source changes and regression test. Review the patch using the workflow you already know.
A reason for every decision
Inspect which checks passed, failed or could not run. Missing labels do not become invented quality claims.
A boundary around each attempt
The declared contract constrains edits and commands. Candidate execution is isolated, with runtime provenance recorded.
An approval that stays yours
A verified result is a proposal. Your team decides whether to create a draft PR and whether to release.
Access, without surprises
Explore freely. Run live with an invitation.
No paid plans or checkout in this hackathon build.
Recorded walkthrough
$0 to explore
Open genuine recordings, follow the candidate edits and inspect the receipts. No account, key or new inference.
Watch a verified repairLive judge workspace
$0 for invited judges
A new repair on the supported workflow, funded by the project’s limited inference credits. Use the access code in the judging instructions.
Try a sample repoWhy live access has limits
Finite hackathon inference credits are reserved for evaluation. Limits help keep live access available throughout judging. Recordings, diffs and receipts remain readable without starting another model call.
If you need an extended evaluation, contact the project owner using the submission testing instructions.
How is this different from a coding assistant?
Autopilot proposes code changes, then an independent verifier evaluates software checks and the ML requirements declared in the repository contract. A patch can pass tests and still be rejected for leakage, serving mismatch or a failed quality guardrail.
Can I see it work without signing up?
Yes. The recordings include actual model activity, candidate diffs and receipts. They are clearly labelled recorded, and viewing them starts no inference. Live judging access uses the code supplied with the submission.
Does VERIFIED mean the model is safe to deploy?
It means the declared checks passed for that snapshot, patch and evaluation data. It is not a guarantee about every future input. Missing labels, unsupported checks and infrastructure failures must remain visible; your team still reviews and ships the change.
Will it change my repository or merge a PR?
Repairs happen in isolated candidate workspaces. You can download a patch or explicitly request a draft PR. Autopilot never merges automatically.
Which repositories work today?
The supported path is Python repositories with explicit commands and an autopilot.yaml contract. Six sample services cover serving parity, preprocessing, leakage, protected-group recall, serialization and numeric boundaries. Arbitrary frameworks and repository layouts are not universally supported.
What happens when a provider fails or a limit is reached?
The interface reports the failure or admission limit without inventing a successful verdict. Existing evidence remains readable. The demo reserves finite inference credits for judging; a brief wait or an owner-assisted live session may be needed.
What can I upload?
Use public code and synthetic or non-sensitive data for the hackathon demonstration. Labels are supplied separately to the verifier. Nebius Sandboxes beta prohibits personal or sensitive files. The hosted demo executes repairs in Nebius Sandboxes.
The next step is an actual repair
See what the patch changed.
Then see what proved it.
Start with the recorded result. Explore the live workspace when you are ready.