DevPulse

🚀 Inspiration

Every engineering team wastes 20–30 minutes daily in standups discussing what's already visible in GitLab. PRs go stale for days. Pipelines fail silently. Issues pile up unassigned. We asked ourselves — what if an AI agent could do all of this automatically?

That's how DevPulse was born.

🧠 What it does

DevPulse is a multi-step AI agent that connects to your team's GitLab repositories and:

Fetches every team member's commits, merge requests, and issues via GitLab MCP Server
Analyzes each contributor's activity and scores them 0–100
Detects blockers — stale PRs (3+ days), failed CI pipelines, unassigned critical issues — each with a priority score 1–10
Generates a human-quality AI standup for every developer using Gemini 2.0 Flash
Takes real action — creates GitLab issues, assigns team members, posts comments autonomously

All of this happens in one click from a dark terminal-style dashboard.

🏗️ How we built it

Agent Architecture — 5-node LangGraph pipeline:

Node 1 — Contribution Analyzer pulls commits/MRs/issues from GitLab REST API
Node 2 — Blocker Detector scans for stale PRs, failed pipelines, unassigned issues
Node 3 — Gemini 2.0 Flash generates personalized standups per developer
Node 4 — Action Executor uses GitLab MCP Server to create issues, assign users, post comments
Node 5 — Gemini writes a final agent summary report for the engineering manager

Tech Stack:

Google Cloud Agent Builder + Gemini 2.0 Flash for reasoning
GitLab MCP Server for real write actions (create/assign/comment)
LangGraph for multi-step orchestration
FastAPI (Python 3.13) backend
React + Vite frontend deployed on GitHub Pages

🚧 Challenges we faced

GitLab MCP authentication — getting the MCP server to authenticate and take write actions required careful token scoping and error handling
Dependency conflicts — Python 3.13 + LangGraph + MCP had conflicting pydantic versions that needed careful resolution
Concurrent async calls — fetching commits, MRs, and pipelines simultaneously across multiple repos without rate limiting
Prompt engineering — making Gemini write standups that sound human, not robotic, required several iterations

🏆 Accomplishments

Built a fully working 5-node LangGraph agent in under 48 hours
GitLab MCP integration that actually writes to GitLab — not just reads
Real-time blocker detection with intelligent priority scoring
AI standups that engineering managers actually want to read
5-person team, each owning an independent module that plugs into one pipeline

📚 What we learned

How to build production-grade multi-step agents with LangGraph
Deep integration with GitLab MCP Server for agentic write actions
How Gemini 2.0 Flash handles structured engineering data for natural language generation
Async Python patterns for concurrent API calls across multiple repos