Skip to content
View PersusUS's full-sized avatar
4 followers · 4 following

Highlights

  • Pro

Block or report PersusUS

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
PersusUS/README.md




Persus

Jesus Perez Bazarot, 21, Seville. I work on the parts of AI that break. I mean to move the world; AI is the lever. Searching for safe AGI.

AI/ML research, continual learning, benchmarks, edge




About

Fourth-year Computer and Electronics Engineering at the Universidad de Sevilla, back from a year of Data Science and AI at the Beijing Institute of Technology. I learn by building, breaking and rebuilding. Long term: a master's in AI at MIT, pointed at something worth keeping.




Stack

Languages: Python, C, C++, Assembly. Deep learning: PyTorch, CUDA, Triton, quantisation. AI/ML: transformers, state space models, RAG, benchmarking, agents. Backend and data: FastAPI, React, PostgreSQL, pgvector, Docker, Neo4j. Spoken: Spanish, English, Chinese, Japanese (work in progress).




Work

2026 to now: in progress. 2025 to 2026: independent AI/ML research on world models, continual learning and LLM benchmarking; AI/ML developer at OrgaAI. 2024 to 2025: co-founder and co-CTO of ByTheWay, carpooling. 2023 to 2024: CTO at NetKey, NFC networking hardware. 2023: speaker at Telefonica innovaTE. Education: BSc Computer and Electronics Engineering, Universidad de Sevilla, 2023 to 2027.




Projects

WMF Benchmark: what a world model forgets when it learns a second taskPerseo: a desktop assistant that listens and watchesHybridMamba-11: 31.8M parameters in 13.6 MB, in OpenAI's Parameter GolfMultilingual: Chinese against Western frontier models, nine languagesNightShift: an autonomous queue that claims one task a nightMAGI: three personas answer, and any one veto sinks the verdictTraces: 15 annotated agent traces, a nine-type error taxonomyKotoba: the whole JLPT, N5 to N1, spaced and offline, in Spanish


All projects




Uptime, public repositories, last push, languages by share, and pushes in the last thirty days - read from the GitHub API once a week.




Now

Organising the AI talks at the Universidad de Sevilla. Research on pushing the TPU state of the art. Building Perseo, the assistant that listens and watches.




Contact

Website Email GitHub LinkedIn Twitter Instagram 



N 42 21 36 W 71 05 31, MIT, Cambridge, Massachusetts. Not there yet. Soon there.

Pinned Loading

  1. WorldModelsBenchmark WorldModelsBenchmark Public

    Component-level benchmark for catastrophic forgetting in world models. Two negative results: forgetting does not follow the labelled task-distance axis, and it happens in the encoder, where the usu…

    Python 1

  2. openai/parameter-golf openai/parameter-golf Public

    Train the smallest LM you can that fits in 16MB. Best model wins!

    Python 5.2k 3.3k

  3. MAGI MAGI Public

    JavaScript

  4. Perseo Perseo Public

    Asistente personal con nucleo siempre encendido: voz en tiempo real por Gemini Live, triaje de correo con modelo local, memoria en Obsidian y confirmacion por niveles. Python + Tauri + React.

    Python 1

  5. selective-scan-cuda selective-scan-cuda Public

    Five CUDA implementations of the selective state-space scan, up to a single-pass decoupled look-back kernel at 79.6-91.5% of peak memory bandwidth

    Cuda

  6. vlm-kv-capacity vlm-kv-capacity Public

    Analytical model: video VLM serving is limited by KV cache capacity, not bandwidth. 113x fewer concurrent users than text on the same node.

    Python