CalCritters
Interactive campus ARG where students scan QR codes to meet LLM-driven characters. Reached 150+ active users.
Machine Learning Engineer
BETODAMIAN
Building and scaling large models and ML systems. Currently developing genomic foundation models at Lawrence Berkeley National Lab.
// autoregressive sample · genomeocean
About
UC Berkeley Data Science
I work on ML systems and large-model scaling — the unglamorous machinery that makes big models train and serve fast: inference pipelines, distributed training, and transformer internals.
Right now that means genomic foundation models at the Joint Genome Institute, where I build generation and evaluation pipelines for autoregressive models that produce DNA one token at a time.
My focus is robust, high-performance inference and training — squeezing throughput out of attention kernels and quantization, and keeping multi-GPU jobs honest on HPC clusters.
Outside research I teach programming, and I tinker with networking and hands-on computing for fun. I'm graduating in December 2026 and open to ML engineering roles.
Experience
Where I've been working
- Developing and evaluating large genomic models like GenomeOcean, an autoregressive transformer that generates DNA token-by-token.
- Built generation and evaluation pipelines using LLM sampling strategies and naturalness scoring across 1M+ sequences, improving F1 by 250%.
- Optimized inference with FlashAttention-2 and 4-bit quantization (+125% throughput); scaled training across 20 GPUs on the Lawrencium HPC cluster.
- Teaching programming and software design to 30+ students across 3 locations, building 60 reusable Python and JavaScript project codebases.
- Developing curriculum modules on foundational coding principles and modern software architecture for beginner learners.
- Providing technical mentorship and debugging support, fostering a collaborative environment for technical growth.
Selected Work
Projects
Engineering scalable ML systems and foundation models at the intersection of computer vision and large language models.
AI-Powered Tutoring Note Assistant
Chrome extension that turns tutor notes into structured session summaries via the Gemini API. Saved ~10 min/session for 20+ tutors.
Real-Time ONNX Hand Tracking
Real-time computer-vision system tracking 21 hand landmarks at 120 FPS — with no MediaPipe dependency.
Zonez
Web app that aggregates student availability across time zones to visualize overlapping study windows. Helped 50+ students coordinate.
Toolkit
Key skills
Languages
- Python
- C++
- C
- Java
- SQL
- JavaScript
ML & Modeling
- PyTorch
- Hugging Face Transformers
- ONNX Runtime
- NumPy · Pandas
- Transformer Architectures
- Quantization (GPTQ/AWQ/GGUF)
DevOps
- AWS · Azure
- HPC Clusters (SLURM)
- CUDA · Linux/SSH
- Git · REST APIs
- Node.js · React