I build coding agents, evaluation infrastructure, ML systems, and performance-sensitive developer tools.
Most recently a Software Engineer Intern on the Next.js team at Vercel, where I built an AI maintainer for Next.js and used it to close 1,500 GitHub issues in one month.
- Built an AI maintainer for Next.js that investigates GitHub and customer reports end to end, from triage and reproduction through canary verification, regression bisection, tests, and fixes.
- Ran it as a human-reviewed queue that closed 1,500 GitHub issues in one month, taking the open backlog from 2,244 to 995 with 3 reopens, and wrote the Next.js blog post on the campaign.
- Owned the product and agent stack behind it, including resumable specialist workflows, model routing, cost and evaluation tracing, and an Eve harness benchmark on Terminal-Bench 2.1.
- Shipped Next.js Agent Feedback, a human-in-the-loop path where coding agents draft deidentified reports on the framework friction they hit and the developer edits, sends, or discards each one before anything is submitted.
- Merged 26 Next.js and Turbopack PRs plus upstream SWC and notify-rs fixes, including a 42× HMR invalidation speedup, a React Compiler precheck that cut compiler pipeline time 19.64% on real v0 modules, and a worker-lifecycle fix that eliminated 100% of measured worker leak growth.
- llm-lab: Language-model training and systems laboratory, built from bigrams through Transformers with tokenization, FineWeb-Edu data pipelines, profiling, checkpointing, and multi-device training. Trained a Transformer across 8 TPU v5e devices, processing 39.85B training tokens at 2.63M tokens/s. Rebuilt that baseline into a 2026-era architecture through a cumulative ladder of 18 controlled experiments covering RoPE, GQA, SwiGLU, RMSNorm, MLA, mixture-of-experts, Kimi Delta Attention, and multi-token prediction. Hand-wrote a Muon optimizer that matched
torch.optim.Muonfrom identical initialization. - 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