AutoDevOps AI

An autonomous Gemini 3 agent that performs deep root-cause analysis and produces verified stabilization blueprints and post-mortem reports while avoiding unsafe production changes.

Comment

Inspiration

Traditional CI/CD tools can tell you that something broke, but often stop short of explaining why or how to fix it. I wanted to explore the Gemini 3 Action Era by building a persistent SRE teammate that doesn’t just suggest code, but orchestrates a complete reasoning lifecycle to help stabilize complex codebases.

What it does

AutoDevOps AI is an autonomous reasoning engine that:

  1. Ingests large portions of repositories natively using Gemini 3’s extended context window.
  2. Diagnoses failures using long-horizon reasoning to isolate deep-seated regressions.
  3. Verifies strategies in a secure sandbox to ensure zero side effects.
  4. Synthesizes a high-fidelity Post-Mortem Analysis and stabilization blueprint for human review.

How it was built

The Brain: Powered by Gemini 3 Pro, using Thought Signatures to maintain continuity across multi-step diagnostic and repair loops.

The UI: A high-fidelity real-time dashboard built using AI Studio Build Mode for "vibe-coding" professional interfaces.

The Backend: A FastAPI action server that handles repo ingestion and provides the AI with "eyes" on the terminal logs and test suites.

Challenges I ran into

The biggest challenge I faced was ensuring the agent didn’t lose contextual continuity during complex, multi-file analysis. I addressed this by strictly implementing Thought Signature persistence, allowing the model to self-correct and pivot its strategy based on previous failed attempts.

Accomplishments that I’m proud of

AutoDevOps AI demonstrated strong agent confidence when isolating issues such as null-pointer exceptions and dependency mismatches. Seeing the agent recall a failure from a previous attempt and autonomously adjust its strategy felt like a genuine “Action Era” milestone for me.

What I learned

I learned that with a 1M token context window, reliance on traditional RAG techniques can be significantly reduced for certain classes of problems. Gemini 3 can natively reason over an entire project structure, identifying relationships between distant modules that were previously difficult for AI systems to capture.

What's next for AutoDevOps AI

I plan to explore integrating Gemini Live for real-time, voice-guided post-mortems and to expand the system’s Safety Boundary logic toward controlled, autonomous PR deployments in staging environments.

Built With

  • firebase-auth
  • firebase-firestore
  • gemini-3-pro/flash-api-(@google/genai)
  • googleaistudio
  • react-19
  • tailwind-css
  • typescript
Share this project:

Updates