Software Engineer Intern on the Next.js team at Vercel in San Francisco.

I build coding agents, evaluation infrastructure, ML systems, and performance-sensitive developer tools.

• Vercel and Next.js: Built an AI maintainer for Next.js that investigates GitHub and customer reports from triage and reproduction through regression bisection, tests, and fixes. Also built resumable specialist workflows, model routing, cost and evaluation tracing, and evaluated the Eve harness on Terminal-Bench 2.1.
• Next.js and Turbopack performance: Made browser-driven HMR up to 10% faster and designed edit transactions that cut v0 rebuild work roughly in half during large agent edits.
• llm-lab: Built language models from bigrams through Transformers. Trained across 8 TPU v5e devices, processing 39.85B training tokens at 2.63M tokens/s, and later trained modern variants using RoPE, GQA, SwiGLU, and RMSNorm.
• BareTensor: Built a near-zero-dependency tensor and autograd runtime from scratch in C++, with strided tensors, broadcasting, neural-network operations, dynamic autograd, and Python bindings.
• ChatVault: Built private semantic search for WhatsApp that runs locally in the browser using quantized MiniLM, Rust, and WebAssembly. Finding and reproducing a Turbopack issue while building it led directly to my off-cycle Vercel internship.
• Cogniba: Designed, built, and launched a brain-training product with Next.js and Supabase, growing it to more than 2,000 registered users.

• Codeforces Expert with a peak rating of 1800
• 2x ICPC SWERC participant
• Olympiad in Informatics, 2nd in Madrid and 16th in Spain
• 3x Meta Hacker Cup Top 2,000
• 2x Ada Byron Spanish national finalist

X | LinkedIn | Email