Mfon_Repro

Inspiration

Reproducing great papers is one of the fastest ways to grow as an engineer, but it’s hard to start from a blank repo and messy PDFs. Repro was inspired by the clarity of LeetCode-style tasks and the desire to make “implement the paper” a simple, guided workflow.

## What it does

Repro lets you upload a research paper and instantly get a structured coding challenge: a scaffolded multi‑file project, tests, and an IDE workspace. You implement the missing logic, run tests in a sandbox, and submit; all from the browser.

## How we built it

We built everything on Next.js and deployed on Vercel. The backend lives in API routes with Neon Postgres for state, Vercel Blob for PDF storage, QStash for background processing, and Vercel Sandbox for running submissions. Gemini 3 powers the scaffold generation with structured JSON output. The frontend renders a split‑pane paper viewer + Monaco IDE.

## Challenges we ran into

• Reliable JSON from the LLM; solved with strict schema validation and repair logic.
• Handling large PDFs and async workflows within serverless limits.
• Securely loading uploaded papers while keeping the UX fast.
• Keeping the IDE state consistent across multiple files.

## Accomplishments that we're proud of

• End‑to‑end flow: upload → scaffold → implement → run → results.
• A LeetCode‑style UI that feels production‑ready.
• A resilient LLM pipeline that can recover from malformed output.
• Full serverless deployment with zero custom backend.

## What we learned

We learned how to orchestrate multi‑step AI workflows inside serverless constraints, how to run unsafe code in the cloud and return the result to the user, and how to build a smooth coding experience entirely in the browser.

## What's next for Mfon_Repro

• Richer paper parsing and citation‑linked specs.
• Better test generation and grading.
• Faster code generation.
• More languages (Go, Rust, C++) and faster runtimes in the sandbox.

Built With

• gemini
• neon
• next.js
• postgresql
• upstash
• vercel