Network Monitor

Replay historical network telemetry on a 3D globe. Watch how a machine-learning model scored outage risk over time, and see where outages were actually recorded.

🟢 Historical Replay Available — The prediction API is connected and the demonstration data is ready.

Replay timeline

Progress through the chosen day, outages found, and the numbers behind each city dot.

Awaiting data...
Live inference Try the prediction API Collapsed by default — see why a small model is fast, try a live score, or copy a curl.

Why a small purpose-built model

Scoring outage risk on telemetry is a high-volume job: many circuits, many minutes. A compact model such as XGBoost is trained on your network’s own history, returns a probability in milliseconds, and can run near the data with predictable cost. A general-purpose LLM with prompts is built for language, not for thousands of metric rows per minute — typical hosted chat APIs take seconds per call and bill per token.

The globe above replays scores already stored in the database (GET /predictions). This panel calls the live POST /predict endpoint separately so you can feel the latency of one score.

Terms (precision, recall, PR-AUC, ROC-AUC, leakage)
Precision
When the alarm beeps, how often is there really a fire? High precision = few false alarms.
Recall
Of all the fires, how many did the alarm beep for? High recall = few missed outages.
Precision/recall trade-off
Lowering the threshold usually catches more fires (recall up) but also beeps for burnt toast (precision down).
PR-AUC
Precision-Recall Area Under the Curve. Try every threshold, draw precision vs recall, measure the area. Best single number when positives are rare.
ROC-AUC
Probability that a random positive row scores higher than a random negative. Can look flattering when positives are rare.
Leakage
When the model is accidentally given the answer (or information from the future). Great score, learned nothing useful for the next week.

Send six measurements

Leave advanced fields closed and the API fills the same rolling and lag defaults the curl uses. Reply shape: probability 0–100; is_outage_predicted is true above 80. Public cap: 30 predictions per visitor per minute.

Full schema at /api/docs
Ephemeral demo Try your data One file is network measurements. The other is outage tickets. We match them in time so the globe can show which circuits actually failed.

These files are not saved on the server. They are held in memory for this demo only, then deleted after 30 minutes. They never go onto the public globe.

  • Metrics CSV: maximum 100 MB and 200,000 rows; one calendar month is processed.
  • Incidents CSV: maximum 5 MB and 2,000 rows.

Waiting to validate.