Command center

Tonight's board

Run a bundled scenario, then inspect incidents, anomalies, and the grounded report. Use the model trace, facts vs rewrite, and closed-book Q&A when you walk an interviewer through it.

api checking

Jobs

0

Incidents

0

Anomalies

0

Run a scenario

No jobs yet. Run Checkout cascade for a full report, Healthy baseline for an empty ledger, or Degraded partial for a sparse-data completion.

How a shift starts

  1. 1. Confirm API health is ok — the pill above is live.
  2. 2. Run Checkout cascade first. It is the noisy, production-like path.
  3. 3. Open Incidents. You should see one candidate, not fourteen alerts.
  4. 4. Read facts vs rewrite. Toggle heuristic and Groq. Facts stay put.
  5. 5. Ask CPU vs baseline, then ask about the weather — the second must refuse.

What the numbers mean

Jobs are in-memory analysis runs for this API process. Incidents are correlated candidates. Anomalies are raw detector hits. A healthy window can complete a job with zero incidents — that is a success, not a missing chart. Degraded partial is for missing-signal resilience: the run summary will warn, the board will still render.

Interview walkthrough

What to say while you click

  1. 01 · What this is

    This is an AI systems project, not a trained model. Detectors are z-score and MAD. Correlation and RCA are graph heuristics. The LLM is optional last-mile rewrite.

  2. 02 · Facts vs rewrite

    Left column is immutable detector output. Right column is a narrative over that JSON. Toggle heuristic vs Groq on the same incident — the facts do not move.

  3. 03 · Model trace

    Every run records provider, model, tokens, latency, cache hits, and fallback. If Groq 404s, the board still has a heuristic report instead of a blank page.

  4. 04 · Closed-book Q&A

    Ask what checkout CPU was versus baseline and it cites the pack. Ask about weather or a service that was not in the run and it refuses. That refusal is the product.

After the first run

  1. 01

    Confirm the clock

    API health on :8000, then a bundled scenario. Checkout cascade is the noisy path. Healthy baseline is the empty-ledger control.

  2. 02

    Read one candidate

    Fourteen detector hits should collapse to one incident. If they do not, correlation failed — that is a bug, not a preference.

  3. 03

    Argue with the origin

    Primary service is a ranked guess. Evidence rows are facts. Contain the origin, not every downstream waiter.

  4. 04

    Write the handoff

    Paste the engineering section into the ticket. Facts stay attached to detector ids so the next shift does not re-litigate the window.

  5. 05

    Close the state machine

    reviewed, then approved or rejected with a name and a note. Only approved reports may leave through a webhook.

Operator notes

Artifacts outlive the board

Restarting uvicorn clears jobs. JSON under artifacts/ui/ remains. Re-run a scenario to refill the console.

Retrieval is on for console runs

Jobs pass data/knowledge/runbooks and incidents. Citations appear when snippets overlap the evidence bundle.

No local model

Compose is heuristic over the evidence bundle. Groq or OpenAI is an optional rewriter. The model trace on Overview tells you which one actually ran.

Caches are under artifacts/cache

Identical compose inputs reuse output. Delete the cache directory if you change heuristic copy and need a fresh narrative.

Prometheus is opt-in

Host allowlist, HTTP and private-network flags default off. A file path is enough for every bundled scenario.

Review requires a name

The API returns 400 without a reviewer. Webhooks only fire for approved reports after URL policy checks.