Input
Incident context
02Operational AI Tooling
Incident Triage Copilot
Live incident-triage demo that turns operational context into a validated, structured response with deterministic fallback behavior.
- Python
- FastAPI
- Next.js
- TypeScript
- Cloud Run
- Ownership
- Owned project
- Timeline
- March – April 2026
- Status
- Completed public demo
Incident Triage Copilot
Public demo using a built-in sample; no real incident data.
Overview
A Next.js and FastAPI application that accepts alerts, logs, metrics, deployment notes, service context, and incident descriptions, then produces a structured summary, severity, impacted service, root-cause hypothesis, immediate actions, and confidence.
Problem
Incident responders need fast, consistent triage from messy operational context without depending on unvalidated free-form model output.
My role
Built the typed frontend-to-backend workflow, provider adapters, deterministic heuristic mode, response validation, failure handling, tests, and public deployment.
System architecture
Surface
Next.js input surface
Contract
FastAPI contract
Triage
Triage path
Guard
Validation / fallback
Output
Product response
Deploy
Deployment
What I built
- Structured incident-input form with sample scenarios and input limits.
- Typed Next.js-to-FastAPI API contract.
- Frontend validation before response rendering.
- Deterministic heuristic triage mode.
- Optional Gemini and Groq provider adapters.
- Provider timeout and fallback when credentials are absent or output validation fails.
- Backend pytest coverage, Ruff validation, and a frontend production build.
- Vercel frontend and Google Cloud Run backend.
Technical decisions
- Structured outputs rather than unbounded text.
- Built-in sample incidents so the demo is useful without any real data.
- Optional model providers, so the demo is not provider-dependent.
- A deterministic fallback path that runs with no credentials configured.
- Request-length limits and visible public-data safety guidance.
Testing and validation
- 16 passing backend tests at the current evidence snapshot.
- Ruff validation across the backend.
- Successful frontend production build.
- Current main CI succeeded.
- Live public page verified at incidentcopilottriage.com.
Measured evidence
16
Backend tests passing
Latest successful main-branch CI run 32055680295 (2026-08-17), rechecked 2026-08-18.
Point-in-time count, not a live counter.
6
Structured response fields returned per triage
Counted from the response schema in the public repository.
A working demonstration that a model-backed triage step can be typed, validated, and degraded safely — the response shape holds whether a provider answers, times out, or is not configured at all.
Limitations
- Decision support, not an autonomous incident responder.
- Use sample or sanitized data only.
- No incident history.
- No Slack or PagerDuty integration.
- No team handoff or timeline generation.
- No published accuracy, user-count, or mean-time-to-resolution metric.
Technology stack
- Next.js
- FastAPI
- Python
- TypeScript
- Cloud Run
- Vercel
- pytest
- Ruff
Data and privacy
The interface warns users to submit sample or sanitized input only. No incident data is retained.