Dengi

Inspiration

Just about every application has bugs, but not everyone has the time to debug. From a local restaurant website to a student's personal projects, Dengi is there to make sure bugs go from discovery to resolution faster than ever before.

Developers today lose countless hours manually digging through vague error logs, routing them to the wrong owners, and writing repetitive fixes. We asked ourselves — why shouldn’t AI triage and propose code fixes before a human even opens the terminal? That led us to build Dengi, an autonomous multi-agent system that intelligently classifies, analyzes, and proposes actual git-ready code patches.

How we built it

Built using Google ADA A2A (agent-to-agent) architecture. Created 3 specialist AI agents (Backend, Frontend, Infra) + Dengi Orchestrator. Used Runner-based execution with streaming session memory for continuity. Wrote a strict JSON schema validator so AI outputs are always developer-safe. Integrated GitHub APPs to auto-commit and PR fixes.

Challenges we ran into

Making sure our fixes from our agents went directly published to a testing branch proved as difficult as we expected. We also had to make sure our fixes were good and publishable with good prompt engineering. We were aware of AI possibly hallucinating and we did not want to just push slop and pass them as fixes.

Accomplishments that we're proud of

Dengi recommends real file patch diffs as well as making an educated engineering decision, not just a chat response. 1st time using Google's ADK A2A multi-agents and revolved our work process around it to make our own dev team. Imported Dengi as a docker container into our server to allow it to trace error logs. This also means Dengi has the ability to be entered as a docker container in any server.

What's next for Dengi

A web dashboard to work alongside directly Dengi to see live updates to the fixes, monitor Dengi's thinking process and customize its scope.
On the fly agent creation for every service needed.