Rate a Hack

How This Started

One of the hardest parts of a hackathon isn’t building — it’s figuring out whether your idea is actually original before you sink hours into it. We wanted a fast “sanity check” that could (1) show similar prior projects and (2) give concrete ways to differentiate.

What it does

Rate a Hack takes a team’s hackathon idea and compares it against a database of past hackathon projects collected from Devpost and GitHub. It returns:

• an originality / similarity score
• a list of closest-matching projects (so you can see what already exists)
• a set of idea suggestions to help narrow scope and differentiate the concept

How it works

Data pipeline

1. Harvest project data and normalize it into a JSONL corpus.
2. Convert each project into a vector representation using SentenceTransformers embeddings.
3. Index the vectors with FAISS for fast nearest-neighbor retrieval.

Scoring + feedback loop

When a user submits an idea:

• the backend embeds the idea text
• FAISS retrieves the nearest matches
• cosine similarity is computed against those neighbors
• an originality score is produced (calibrated to emphasize distinguishing concepts vs. superficial wording)
• the system returns:
  • the score
  • closest matches
  • suggestion prompts to make the idea more specific / defensible

Team

Built with:

• jacobwoodworth Woodworth
• Sypitkowski Sypitkowski
• Christopher Kopiwoda (me)

My main contribution

I worked across the data pipeline and product layer of the project:

• Built the project database pipeline, including the web scraping and data structuring that generated the JSONL corpus of past hackathon projects from DevPost.
• Designed and implemented the frontend (React + TypeScript), including the character-based feedback UI, scoring display, tab system (Projects / Idea Suggestions / Customization), and overall user interaction flow.
• Helped connect the backend and frontend, wiring API endpoints to the UI, handling request/response formatting, and ensuring idea submissions flowed cleanly from the interface to the retrieval engine and back.

My focus was on making the system usable end-to-end: reliable data ingestion, a clean interface, and seamless communication between the model and the user experience.

Challenges

• Calibration: Getting a score distribution that isn’t “everything is 80% similar” required tuning weights and focusing on concept-level uniqueness.
• Data collection: Rate limits and staying within platform terms made ingestion slower and more careful than expected.
• User clarity: A single number isn’t enough — it needs examples (nearest matches) and actionable suggestions to be useful.

What we’re proud of

• A complete end-to-end workflow: idea → retrieval → score → UI.
• A reusable JSONL dataset of prior hackathon projects.
• A clean UI with hand-drawn character feedback and multiple modes (projects / suggestions / customization).

What’s next

• Deploy on a dedicated server (not local-only).
• Expand the dataset toward hundreds of thousands of projects (ideally all of Devpost).
• Make the score more explainable (e.g., highlight which concepts/features contributed most to similarity).

Media & Documentation

UI Screenshots

1 / 4

Tools Used