Devmate
AI that connects builders for hackathon
Created on 8th November 2025
•
Devmate
AI that connects builders for hackathon
The problem Devmate solves
The Problem It Solves
Finding teammates for hackathons is slow and unstructured. Builders often rely on chats or forms that fail to match people by actual skills, goals, or timezones.
Devmate automates this using an AI-powered search pipeline that understands intent and returns ranked, explainable teammate matches.
What It Does
Devmate processes each search through three coordinated agents:
- Planner interprets natural language requests and extracts structured intent.
- Executor embeds the query, performs vector search on Postgres with pgvector, and applies filters like role, timezone, and availability.
- Evaluator ranks results using a blend of similarity scores and LLM-based reasoning.
It returns transparent, structured recommendations showing how and why each builder was selected.
view architecture details here : devmate-architecture.md
How It's Built
Devmate is built with Next.js 16 and React 19 for the frontend, powered by Node.js and Express on the backend.
The system integrates the Vercel AI SDK for orchestrating calls to OpenRouter models (GPT-4.1 and GPT-4.1 mini).
Embeddings are generated and stored using pgvector in PostgreSQL, managed via Drizzle ORM.
Each search follows a clear AI pipeline:
- The Planner Agent (GPT-4.1 mini) uses the AI SDK to parse the free-form request into a structured plan.
- The Executor Agent retrieves matching profiles by running vector and metadata queries on Postgres.
- The Evaluator Agent (GPT-4.1) re-scores top results and generates reasoning explanations.
- The Orchestrator combines all outputs, logs reasoning and metrics, and returns a ranked JSON response to the UI.
The frontend renders the ranked teammate cards and reasoning trace in real time using server actions and the AI SDK streaming API.
Why It Helps
- Understands real intent beyond keywords
- Explains every result transparently
- Runs quickly with pre-embedded builder profiles
- Integrates easily into hackathon or builder platforms for AI-assisted teammate discovery
Challenges I ran into
Challenges I Ran Into
-
Getting the three agents to communicate properly was difficult at first. The planner’s output wasn’t always structured enough for the executor, which caused missing or empty results. Enforcing strict Zod validation and typed payloads between phases fixed the flow.
-
Embedding generation also failed midway due to API rate limits and timeouts. I added batching and retry logic to make it consistent.
-
Vector queries initially returned no results because of Postgres connection limits, which I resolved by tuning the pool size and query handling in Drizzle.
-
Deployment wasn’t completed due to time constraints and setup complexity with database extensions and environment configuration, so the final demo was run locally.
PS : What's Pending / Improvements
- Adding a profile view with contact options would make results actionable.
- Integrating with Devfolio through an API or embed would enable direct use inside hackathons.
- Deployment is still local; setting it up with proper environment config and scaling is next.
Technologies used
Cheer Project
Cheering for a project means supporting a project you like with as little as 0.0025 ETH. Right now, you can Cheer using ETH on Arbitrum, Optimism and Base.