Building IncidentCopilot: Establishing a Local-First AI DevOps Development Foundation

Project: IncidentCopilot — AI DevOps Incident Investigation

Milestone: 1 — Repository & Local Development Foundation

Status: ✅ Completed

#

🎯 Why Start With the Foundation?

When building an AI-powered DevOps system, it’s tempting to jump straight into the LLM.

For IncidentCopilot, I deliberately chose not to.
Evidence first. AI second. Human in the loop.

The AI should reason over verified evidence, not replace deterministic systems like parsing, normalization, persistence, or correlation.

So Milestone 1 focused on:

• Repository structure
• Local dev environment
• Backend & frontend foundations
• Config management
• Testing setup
• Docker & Compose
• Documentation & reproducibility

#

🖥️ Local-First Decision

IncidentCopilot is intentionally local-first. No reliance on:

• AWS / Azure / GCP
• Paid APIs
• Proprietary SaaS infrastructure

Instead, the stack runs via Docker Compose:

• FastAPI
• PostgreSQL
• Qdrant
• Ollama
• React

#

📂 Repository Structure

Backend packages were defined but left intentionally empty — establishing architectural direction without premature implementation.

#

⚙️ Backend Foundation

FastAPI app with two endpoints:

Config management viapydantic-settings #

Testing with Pytest + FastAPI’sTestClient #

Dockerized backend (minimal container, no DB/AI yet)

#

🎨 Frontend Foundation

• React + TypeScript + Vite + Tailwind CSS + Lucide icons

• Minimal shell: IncidentCopilot — AI DevOps Incident Investigation

• Cleaned unused Vite starter files

Dockerized frontend with Node-based build image

#

🛠️ Real Problems & Fixes

Node.js mismatch: upgraded from v20 → v24 for Vite #

Docker Desktop: CLI installed but engine not running — fixed by starting Docker Desktop #

Windows make: used mingw32-make instead of GNUmake #

Git hygiene: fixed invalid UTF-8 README + refined.gitignore

#

✅ Verification

• Git hygiene → clean
• Backend tests → 1 passed
• Frontend lint → 0 errors
• Frontend build → ✓ built
• Docker Compose config → valid
• Backend & frontend containers → running locally

#

🧩 What We Didn’t Build (Yet) Milestone 1 deliberately excluded:

• PostgreSQL models
• Log ingestion APIs
• Parsers (Nginx, Kubernetes, Docker, GitHub Actions)
• Normalization & correlation
• Qdrant + RAG integration
• Ollama integration
• Structured AI diagnosis
• Full incident dashboard

These belong to future milestones.

#

🏗️ Architecture Principle

The LLM will sit after the deterministic evidence pipeline.

#

📌 Key Takeaways

1. Foundation work = real development
2. Verification > assumptions
3. Starter templates should be questioned
4. Local-first changes dev strategy
5. AI shouldn’t be the first thing we build

#

🔮 What’s Next?

Milestone 2 — FastAPI Foundation + PostgreSQL

Moving toward:

#

🏁 Final Thoughts

IncidentCopilot is still at the beginning. No AI diagnosis yet. No RAG. No ingestion pipeline.

And that’s okay.

Milestone 1 established the engineering environment needed to build those capabilities correctly.

The project now has:

• Structured monorepo
• FastAPI + React/TypeScript
• Tailwind CSS
• Config management
• Testing
• Docker + Compose
• Verified local workflow

Most importantly:

Build the evidence pipeline first. Let AI reason over verified evidence later.

🔗 GitHub: https://www.github.com/richardatodo/incidentcopilot ➡️ Next: Milestone 2 — FastAPI + PostgreSQL

source & further reading

dev.to — original article I benchmarked 4 models on the bash macOS still ships - and my harness failed first How to Stop Runaway LLM Agent Loops from Draining Your Credit Card Designing browser music controls around the next beat