IncidentCourt

The problem

When a service incident unfolds, engineers have to compare metrics, logs, and traces while deciding which explanation deserves further investigation. A plausible AI-generated diagnosis can sound confident even when the relevant observation is missing, another explanation fits, or the evidence is incomplete. Responders need a way to examine the basis for an assessment, not just its final sentence.

What we built

IncidentCourt is an investigation workspace that turns telemetry into an evidence trail. It compares baseline and later observations, proposes an initial service hypothesis, challenges it with independent signals, and presents supporting observations, counterevidence, alternatives, and uncertainty. The reviewer can inspect evidence rows, see the SQL query behind each observation, explore observed service dependencies, add separate review notes, and export the report.

The product is designed for engineers investigating service incidents. It identifies candidates for further investigation; it does not assert a proven root cause or take remediation actions. A human remains responsible for interpreting the evidence and deciding what to check next.

How Nebius and NVIDIA fit in

IncidentCourt calls NVIDIA Nemotron 3 Super (nvidia/nemotron-3-super-120b-a12b) through the Nebius Token Factory API. The planner selects an initial hypothesis from supplied services and evidence. The critic requests a follow-up check, and a third request produces an explanation referencing the collected evidence. The application checks the requested and returned model identity, validates structured outputs and evidence identifiers, and records provider-call receipts. It fails an invalid response instead of silently falling back to a different model or an offline answer.

In one browser-started run using sampled telemetry, the three Token Factory requests returned from the selected model. The resulting report contained 21 evidence records and 14 queries; checked evidence and query references resolved. This verifies one integrated workflow, not diagnostic accuracy or the factual correctness of every generated sentence. Token usage and provider responses are not an independent billing record.

Nebius serves the model. IncidentCourt itself performs the bounded telemetry queries and manages the evidence and review workflow. The demonstration uses a small, disclosed sample from controlled RCAEval Online Boutique experiments, not production incidents. The app also includes the local Meterless H-MEM reference for recalling eligible observations; previous findings are treated as leads and rechecked against current telemetry. H-MEM returned no prior matches in the documented Nebius run.

What we learned and built with

The application uses Python, JavaScript, SQLite, Node.js, NVIDIA Nemotron, Nebius Token Factory, the Meterless H-MEM reference, and RCAEval sample data. Codex helped build and test the provider integration, citation validation and fail-closed behavior, and the browser workflow. The code asks for structured, evidence-referenced responses; a valid identifier confirms a reference exists, not that the interpretation is true.

The offline sample uses deterministic rules without a model request. The optional Nebius mode requires explicit configuration and inference opt-in.

Challenges and limitations

Sampled telemetry can omit brief events or present unequal coverage across signals. A dependency edge or correlation does not establish causation. The report may show competing explanations or insufficient data, and reviewers must assess model-generated prose. A single successful run does not establish production readiness, model accuracy, comparative performance, latency, cost savings, or deployment at scale.

Built With

• javascript
• meterless-h-mem
• nebius-token-factory
• node.js
• nvidia-nemotron
• python
• rcaeval
• sqlite