Builder
AI coding agent
Reads the repository, forms a hypothesis, and proposes a patch you review as a diff before it lands.
Cannot decidewhether its own patch is correct.
2 engineering missions live — find your starting point
The Engineering System Behind AI
Build, debug, and ship real AI systems through hands-on engineering missions.
A real repository, mid-incident.
You open the workspace at the moment monitoring fired. The service is healthy, the tests are green, and the model is wrong. Find out why.
app/preprocess.pypython
- keys = sorted(payload.keys())
- return [float(payload[key]) for key in keys]
+ return [float(payload[key]) for key in FEATURE_SCHEMA]
The thing that writes the code does not get to grade it.
Three separate authorities run every mission. Keeping them apart is the whole reason this is not a chat window with a repository attached.
Every mission becomes evidence.
Finishing does not award a badge. The mission records what it was able to verify about you, and stays explicit about what it could not.
This is the record Incident 001 writes. Three skills are backed by the run; the fourth is named as unproven rather than quietly left out.
Evidence lives in the current browser session. Carrying it between sessions is on the roadmap, not shipped.
incident 001 · evidencefraud-service
- Train-serving consistencyVERIFIED
- DebuggingSTRONG
- TestingGOOD
- Input validationNEEDS WORK
Recorded after 7 of 7 deterministic checks passed and the mentor question was answered.
What exists, and what does not.
Available today
2 incident missions, a deterministic verifier, a mentor gate that stays shut until the checks pass, and a three-question starting diagnostic. All of it runs in the browser with nothing to install.
Coming later
A skill graph that persists between sessions, review and build missions alongside incidents, real repository work, and a wider library. None of it is wired yet, and nothing above depends on it.