Nebius × NVIDIA Global AI Hackathon · Best Apps and Agents
Root-cause analysis you can check.
Traceback reconstructs what happened in a failed multi-service request, names the most likely root cause, cites the exact records behind every claim, and then replays the incident with the suspected cause switched off to see whether the failure disappears.
Three systems, one scoreboard
How it works
Deterministic analysis first. The language model only ranks, explains and checks hypotheses against evidence, and it is never trusted without a replay. Click a stage to see it run on a real incident.
What the LLM can and can’t do
- It sees ranked candidates and the records behind them, not the raw firehose.
- Every claim must cite record IDs. IDs that don’t exist are rejected and the model is asked again.
- It can overrule the analyzer, and sometimes wrongly; the replay verifier catches some of that.
Honest limitations
- Evaluation data comes from a simulator: it shows the pipeline handles the modeled failures, not that it works on production telemetry.
- 26 test incidents: a one-incident gap in accuracy is not significant.
- The verifier needs a replayable environment (a canary or feature-flag rollback in production).
- Model answers shown here are recorded from real runs. The static site never calls a model;
make liveruns the simulator and pipeline locally and can optionally call the model with a time limit.