A little about me
I’m Jesús Pérez Bazarot: 21, from Seville, in my fourth year of Computer Engineering at the Universidad de Sevilla after a third year at the Beijing Institute of Technology. I don’t really believe something works until I’ve taken it apart. That has pointed me at very different things: a GPU kernel measured against the bandwidth of the chip it runs on, a benchmark that ended up refuting the method I proposed in it, a startup’s entire stack when I was its only engineer. The habit behind them is always the same: write down what I expect, measure, and say so when I was wrong.
I started early and in the open: CTO of a small hardware startup in my first year of university, a talk on Telefónica’s innovaTE stage, where I also introduced the Mayor of Seville, and a carpooling company I co-founded that didn’t survive. I kept the lessons from all three. I start more projects than is wise, and I finish the ones that matter. Three or four evenings a week I’m on a mat doing martial arts, which is where I learned to lose without quitting.
What I do
I write down what I expect before I measure it. When the numbers disagree, the numbers win, even when the idea they sink is mine.
Right now that means two lines of work. The first is continual learning in world models: a 375-run benchmark I wrote up alone, under review at a NeurIPS 2026 workshop, whose two findings are both negative. Task distance doesn’t predict how much a model forgets, and the forgetting that matters happens in the encoder, where the usual metrics don’t look. One of those results refutes the method I set out to propose, and it went in the paper anyway. The second is the level below the model: hand-written CUDA scan kernels that sustain up to 91.5% of a laptop GPU’s memory bandwidth, and before that a Mamba–Transformer hybrid for OpenAI’s Parameter Golf, made compilable by rewriting its recurrence as a parallel scan in plain PyTorch.
Almost everything else I build, I build to use. Perseo, the assistant I’ve been rebuilding since 2023, reads my mail and calendar and only interrupts me when it matters; NightShift works through a queue of coding tasks while I sleep; Kotoba is how I study Japanese. Next comes a PhD in machine learning, with MIT as the target, and further out, work that points machine learning at something worth conserving. The shark on this site isn’t decoration.
Off the screen
Experience
-
Independent AI/ML Research World models, continual learning, LLM benchmarking2025–Present
-
AI/ML Developer OrgaAI: audio & transcription pipeline, under NDAJun–Sep 2025
-
Co-founder & Co-CTO ByTheWay Carpool (closed)2024–2025
-
Research Intern Minerva Research Group, Universidad de Sevilla2023–2024
-
Chief Technology Officer NetKey: NFC networking hardware, closed2023–2024
-
Technology Speaker, innovaTE Telefónica: NB-IoT & sustainability track2023
Education
-
BSc Computer Engineering Universidad de Sevilla, Computer Engineering specialisation2023–06/2028 (expected)
-
Erasmus exchange Beijing Institute of Technology: CS & Big Data AnalysisSep 2025–May 2026
Selected work
-
WMF Benchmark Catastrophic forgetting in world models: under review at CL4FMAgents @ NeurIPS 2026 (non-archival workshop)2026
-
Selective-scan CUDA kernels Five hand-written scan kernels, up to 91.5% of peak DRAM bandwidth2026
-
HybridMamba-11 OpenAI Parameter Golf, submitted April 20262026
-
Perseo Personal AI assistant: Tauri 2, Rust, Gemini Live2023–Present
-
Agentic Reasoning Traces 15 annotated traces, 142 steps, 9-type error taxonomy2026
-
ClassTranscriber 7-stage lecture pipeline: diarisation, Whisper, local cleanup2025
Achievements
-
Wanted Entrepreneurship Competition First place out of 20 teams2023
-
Debate Competition finalist Provincial level2023
-
NASA Space Apps Challenge Swarm systems and robotics challenges2023
-
Algorithmics Club Universidad de Sevilla, member2023–Present
Toolkit
-
Languages Python, TypeScript, Java, C, C++, Rust, SQL, Assembly
-
Deep learning PyTorch, CUDA, Triton, torch.compile, BF16, quantisation, DDP
-
AI/ML Transformers, state space models, RAG, speech pipelines, diarisation, benchmarking
-
Backend & data FastAPI, React, Supabase, PostgreSQL, pgvector, Docker, Neo4j, WebSockets
-
Spoken Spanish (native), English (C1, Trinity 2020), Chinese (conversational), Japanese (learning)
Where I have been
From a lecture hall in Seville to a campus in Beijing, from a Telefónica stage to a NeurIPS workshop: the places, teams and programmes behind the work above.
