Building intelligent robots that see and act.

Training a VLM for robot understanding — multi-GPU pipeline with Hugging Face models, designed for real-world robotic perception and instruction following.

From perception to manipulation — a Vision-Language-Action model that converts visual understanding into precise robot actions for real-world manipulation tasks.

Real-time ML inference using asynchronous WebSocket architecture — streaming multimodal data (images, tensors, text) for low-latency robot control.

Rewrote the async ML inference server from Python to Rust for performance-critical robotics applications — with Python bindings for interoperability.

Implementation of RT-1 architecture for robotic manipulation — transformer-based policy for real-world task execution from demonstrations.

Scripts and pipelines for fine-tuning large language models with LoRA and QLoRA on custom datasets.