Co-Op Compass

A full-stack co-op application tracking platform built for Drexel students — extended with an AI agent layer that autonomously finds, scores, and tailors co-op opportunities.

Track applications, visualize pipeline progress, analyze trends, and review a ranked shortlist of AI-sourced roles — with a human deciding on every one.

Screenshots

What it does

Students add their co-op applications and track them through a kanban pipeline — Applied, Interview, Waitlist, Offer, Rejected. The analytics page visualizes application trends and status breakdowns over time.

On top of this, an AI agent pipeline (in /agents) autonomously discovers tech co-op postings, scores each for fit, tailors a resume recommendation and outreach draft for strong matches, and surfaces them in an "AI Suggested" view. The user reviews each one and either tracks it as an application or rejects it — the agents find and prepare, the human decides.

Tech Stack

Application

• Next.js 16 + React 19 + TypeScript
• Tailwind CSS
• Supabase (PostgreSQL)
• Recharts
• Vercel (deployment)

AI agent layer (/agents, Python)

• CrewAI — multi-agent orchestration (resume tailoring + outreach)
• Claude API (Haiku) — job scoring, tailoring, outreach drafting
• Pydantic — structured-output validation
• SimplifyJobs listings — job source
• Supabase — shared backend between the pipeline and the app

Features

Tracker

• Kanban pipeline with real-time status updates
• Add applications with duplicate detection via external job IDs
• Search and filter applications by status
• Analytics dashboard with area chart and status breakdown
• Recent activity and upcoming deadlines on dashboard

AI agent layer

• Autonomous job discovery from cycle-relevant tech internship listings
• LLM scoring (1-10) with reasoning and automatic red-flag detection
• Hard filters split from a "verify before applying" checklist — the model gates on what it can judge (role, cycle, seniority), and flags what it can't (pay, visa) for human review
• Content-driven resume selection across multiple resume versions
• A sequential CrewAI crew that suggests specific resume edits and drafts a tailored outreach message for strong matches
• One-click "Track as Applied" graduates an AI suggestion into the real tracker

Design decisions

A few choices worth calling out:

• Workflow, not autonomous agents. The pipeline is a prompt-chaining workflow with a gating step — predictable and debuggable — rather than a model deciding its own control flow. Autonomy was added only where the problem genuinely called for it.
• Structured output over prompt-and-parse. Every LLM call uses forced tool-use validated by Pydantic, with CHECK constraints in Postgres as a second layer.
• Resumes are data, not code. The valid resume set is derived at runtime from a folder of markdown files; adding a resume needs no code change.
• Model chosen by benchmark. Haiku was selected over Sonnet after a side-by-side comparison on real jobs showed near-identical scoring at ~3x lower cost.
• Cost engineering. Result caching, prompt caching, and gating the expensive tailoring crew to only strong matches keep runs cheap.

How to Run

App

1. Clone the repo
2. Create .env.local:

   NEXT_PUBLIC_SUPABASE_URL=your_url
   NEXT_PUBLIC_SUPABASE_ANON_KEY=your_key
   

3. npm install
4. npm run dev

AI agent pipeline

1. cd agents
2. Create a virtual environment and install dependencies:

   uv venv --python 3.12
   source .venv/bin/activate
   uv pip install -r requirements.txt
   

3. Create agents/.env:

   ANTHROPIC_API_KEY=your_key
   ANTHROPIC_MODEL=claude-haiku-4-5
   SUPABASE_URL=your_url
   SUPABASE_KEY=your_key
   

4. Add your resume versions as markdown files in agents/resumes/ (e.g. general.md, ai.md, systems.md)
5. Run the pipeline: python pipeline.py

Scored suggestions are written to Supabase and appear in the app's "AI Suggested" view for review.