Team Claude
Team consisting of NYU CS and Data Science master's students, featuring an ex-PepsiCo data scientist and a senior software engineer with 4+ years of industry experience.
YouTube Video
Project Description
We built Hackathon Judge AI :
A tool that takes the pain out of hackathon judging. Instead of tired human judges rushing through 50+ demos, organizers just submit each team’s GitHub repo, pitch deck, and demo video. Our system does the rest.
We have 3 AI agents running in parallel using Claude LLM via the Anthropic SDK through CrewAI:
Code Agent : clones the repo, checks code quality, architecture, commit patterns, red flags
Pitch Agent : reads the pptx slide by slide, evaluates business model, flags unsubstantiated claims
Demo Agent : uploads the video to Gemini 2.5 Flash, checks if the product actually works or its a fake demo
Once all 3 are done, a Synthesizer agent (also Claude) cross-references everything. Like if the pitch says “HIPAA compliant” but the code has zero encryption it catches that. It generates scores, tough judge questions with evidence, and a voice script ready to read out loud.
We use Redis caching to store intermediate agent results and speed up repeated lookups so the pipeline doesn’t redo work unnecessarily.
The UI is not a chatbot. There’s no chat window anywhere. Instead you see a live pipeline visualization data flows left to right through animated SVG paths, agents light up as they work, checkmarks pop in when they finish. You’re watching AI think in real time through SSE streaming. The pipeline IS the interface.
Tech stack:
AI/LLM: Anthropic Claude Sonnet 4 (via Claude SDK + CrewAI), Google Gemini 2.5 Flash (video analysis)
Backend: Python, FastAPI, CrewAI, Redis (caching), GitPython, python-pptx
Frontend: React, Vite, Tailwind CSS, Framer Motion
Streaming: Server-Sent Events for real-time pipeline updates