About Me
I'm an Applied Scientist at Amazon AGI working on evaluation, recursive self-improvement (RSI) and RL post-training of LLM agents. I build evaluations that expose where agents fail, generate the data that fixes those failures, and increasingly let models run that loop themselves.
Before Amazon, I did my graduate studies at ETH Zurich and UC Berkeley, where I was a researcher in the Sky Computing Lab with Ion Stoica and Joey Gonzalez. My CV is here (PDF).
News
- Nov 2026
Placed 5th of 2,000+ teams in Meta's Open Kernel Challenge with an RSI framework I built, with no prior CUDA experience.
- Nov 2026
Released BOBA-Bench: coding agents solve 95% of tasks alone but only 51% when tasks are chained.
- Mar 2026
Sierra released τ³-bench with the fixes from τ²-Bench-Verified, crediting SABER in its blog post.
- Jan 2026
vAttention and vCache accepted at ICLR 2026, and SABER at an ICLR 2026 workshop.
- Dec 2025
Amazon Nova 2 launched. I contributed RL environments and SFT tool-use data to Nova 2 Lite.
- Jul 2025
Joined Amazon AGI as an Applied Scientist.
- May 2025
HashAttention accepted at ICML 2025.
- Feb 2025
Released The Danger of Overthinking, later covered by IEEE Spectrum.
- Jan 2025
Showed that OpenAI's o1 solves 30% of SWE-bench Verified, not the reported 48.9% (thread), covered by The Decoder. JudgeBench accepted at ICLR 2025.
Experience
- Built evals exposing agent limits, showing coding agents solve 95% of tasks alone but 51% chained, published as BOBA-Bench.
- Built RSI pipelines whose tasks lifted Qwen3.6-35B-A3B above Opus 4.8 on Vending-Bench, an economic agent benchmark.
- Ran a 200-agent factory that turned 41K repos into 7,000+ 20-step coding chains that teach models not to regress prior work.
- Built an agentic audit that scanned 378 RL environments (55K+ tasks) for reward hacks and task correctness.
- Led Amazon AGI's first asynchronous, long-context agentic RL post-training of a 300B+ MoE model on hundreds of GPUs.
- Found errors in τ²-bench (used by OpenAI, Anthropic, Google, Meta) and built τ²-Bench-Verified, whose fixes are in τ³-bench.
- Contributed RL environments and thousands of SFT tool-use trajectories to Nova 2 Lite, matching Sonnet 4.5 on τ²-Bench.
- Built a framework that verifies agents' state-changing actions and cleans stale context, raising task success up to 28% (SABER).
- Showed in an extensive study of o1-style models as coding agents that they overthink instead of acting (paper, 180+ citations), findings adopted by Meta, Microsoft, CMU, Berkeley and others and covered by The Decoder and IEEE Spectrum.
- Co-created learned token hashing for long-context agents (32× fewer tokens) and led its Llama training (HashAttention, ICML).
- Developed error-guaranteed sparse attention and LLM caching for agent inference, published at ICLR as vAttention and vCache.
- Built a coding split of right and wrong solutions that strong LLM judges barely tell apart, published as JudgeBench at ICLR.
Publications
Full list on Google Scholar. Citation counts as of Nov 2026.
Projects
- Built an RSI framework that writes and optimizes GPU kernels and placed 5th with no prior CUDA experience.
- Turned it into a GPU-kernel training environment that teaches models to write fast kernels out of the box.
- Built the multi-agent setting and o1 support (native function calling, reasoning effort).
Education
- 2024 – 2025UC Berkeley, Graduate studies, Computer Science
- 2022 – 2024ETH Zurich, M.Sc. Computer Science, Secure and Reliable Systems
- 2018 – 2022Universidad Politécnica de Madrid, B.Eng. Computer Engineering, top 5% of class
Honors & Awards
- Over 800 citations on Google Scholar in the last two years, and interviewed by IEEE Spectrum (2025).
- Invited reviewer for NeurIPS 2026, ICLR 2026 and ICLR 2027.
- Member of IEEE-Eta Kappa Nu (HKN) and founding president of Spain's first HKN chapter.
- Excellence scholarships from Fundación Mutua Madrileña and the Government of Madrid.