🛡️ OpsPilot

Autonomous Multi-Agent SRE Command Center

From production incident to confidence-scored root cause in under five minutes.

Author — Reyas Khan

---

The problem

When production breaks, an on-call engineer manually pivots across metrics, logs, deployments, and incident history under pressure — a slow, tribal, error-prone scramble while every minute costs revenue and trust.

What OpsPilot does

OpsPilot turns that scramble into an autonomous, evidence-backed investigation. A Commander agent dispatches a team of specialized AI agents that investigate in parallel, correlate their evidence into a causal timeline, determine a confidence-scored root cause, and propose ranked, executable remediations — all streamed live to an Azure-Portal-style command centre over Server-Sent Events.

	Traditional on-call	OpsPilot
Time to root cause	30–60+ min	< 5 min
Evidence gathering	Serial, manual	Parallel, autonomous
Root cause	Tribal guess	Confidence-scored hypothesis
Remediation	Ad hoc	Ranked, executable fixes

---

Demo

📹 Product walkthrough — a full incident, from detection to resolution.

▶️ Watch the full demo on YouTube

---

The differentiator: confidence-gated reasoning escalation

After root-cause analysis, OpsPilot computes a combined confidence score. Only when it falls below threshold (default 70) does it route the full context to the o4-mini reasoning model, which re-examines the evidence from first principles and returns a refined root cause plus a complete reasoning trace. Frontier reasoning stays off the hot path unless it's actually needed — governing cost and latency by design. This is intelligent model routing, not brute force.

Agent	Role	Model
Commander	Orchestration & synthesis	GPT-4o
Metrics / Logs / Deployment	Parallel evidence gathering	GPT-4o-mini
Time Machine	Causal event timeline	GPT-4o
Root Cause	Confidence-scored hypothesis	GPT-4o
Deep Reasoning	First-principles re-analysis (escalation)	o4-mini
Recommendation	Ranked, executable remediations	GPT-4o-mini

Agents are routed by role to configurable Azure AI Foundry deployments and orchestrated as a compiled LangGraph StateGraph with conditional escalation edges. A provider abstraction runs the whole system in zero-credential mock mode, then flips a single env var for live Foundry inference.

---

Run it (for judges)

OpsPilot runs in zero-credential mock mode by default — no Azure account needed to evaluate the full experience. Every path opens the app at http://localhost:3000.

First, from the repository root: cp .env.example .env The defaults in .env.example select mock mode, so this just works.

Option A — Prebuilt images from GHCR (fastest, no build)

cp .env.example .env
docker compose -f docker-compose.ghcr.yml up

Or pull the images directly:

docker pull ghcr.io/rewyekha/opspilot-backend:latest
docker pull ghcr.io/rewyekha/opspilot-frontend:latest

Option B — Build from source with Docker

cp .env.example .env
docker compose -f docker-compose.yml up --build

Option C — Local dev (hot reload)

cd backend && python -m uvicorn app.main:app --reload --port 8000  # mock mode, no credentials
cd frontend && npm install && npm run dev

Bring your own Azure AI Foundry credentials (live mode)

Edit .env:

EXECUTION_MODE=foundry
FOUNDRY_ENDPOINT=https://<your-resource>.openai.azure.com/
FOUNDRY_API_KEY=<your-key>            # or leave empty to use managed identity
COMMANDER_MODEL_DEPLOYMENT=gpt-4o
SPECIALIST_MODEL_DEPLOYMENT=gpt-4o-mini
REASONING_MODEL_DEPLOYMENT=o4-mini

Then re-run any option above. See backend/.env.example for the full variable list (all optional except FOUNDRY_ENDPOINT in live mode).

---

Container images

Image	Registry path
Backend (FastAPI + LangGraph)	ghcr.io/rewyekha/opspilot-backend
Frontend (React + nginx)	ghcr.io/rewyekha/opspilot-frontend

Built and published on every push to main by .github/workflows/docker-ghcr.yml using the built-in GITHUB_TOKEN — no external secrets. Tags: latest, the short commit SHA, and semantic-version tags on v* releases.

---

Repository layout

.
├── backend/            # FastAPI + LangGraph agent runtime (Python 3.12)
├── frontend/           # React + TypeScript + Fluent UI v9 (nginx in prod)
├── infra/              # Azure Bicep infrastructure-as-code
├── evaluation/         # Foundry evaluation datasets and runners
├── demo-workloads/     # Sample apps used to generate real incidents
├── docs/               # architecture diagram assets
├── docker-compose*.yml # local / prebuilt run configurations
└── .github/workflows/  # CI/CD (Docker → GHCR)

Tech stack

React 18 · TypeScript · Fluent UI v9 · FastAPI · Python 3.12 · LangGraph · Azure AI Foundry · Azure OpenAI · Docker · GitHub Actions

---

Built for the Microsoft Agents League Hackathon · Reasoning Agents Track