I’m the technical co-founder of BETTORCA, where I’ve served as the primary engineer behind the company’s product. I started as a full-stack developer, but as the product evolved and we learned more from the market, I became increasingly interested in AI/ML research, particularly how to train models to make predictions that create real value for users.
Outside of engineering, I’m usually in the gym, watching sports, looking for vintage pieces, planning a trip, or working on my golf game.
Active contributor to open-source tools for model training, exploring different approaches through controlled experiments.
LitGPT
Improved multi-turn fine-tuning by focusing training on assistant replies across complete conversations while preserving existing training behavior. Controlled experiments evaluated effects on response quality, context retention, and general capabilities.
torchtitan
Added per-depth loss normalization for Mixture-of-Depths training in DeepSeek models to improve consistency during distributed optimization. Analysis covered gradient scaling, padding, and variable-length sequences to identify when normalization helps.
Python
CUDA
Go
Rust
Bash
SQL
PyTorch
XGBoost
Hugging FaceTransformers, Tokenizers
FastAPIPyTorch DistributedDTensor, DeviceMesh
Lightning Fabric
Docker
Kubernetes
ONNX
MLflow
GitHub Actions
Railway
PostgreSQL
Qdrant
Redis
pandas
PyArrow