At InkyLab Technologies, I deployed open-weight LLMs on vLLM, raising throughput 150% and cutting GPU cost 40%. I also profiled PyTorch inference across 16 GPUs and optimized C++ and CUDA bottlenecks, lifting throughput 60% and lowering latency 35%.
At AIRE, I cut contract-chat response latency from 2.5 seconds to 800 milliseconds with vLLM serving, quantization and caching. Across my work, I’ve built RAG systems, AI agents and computer vision pipelines, including an agentic investigation platform that monitors 500 sources and 10,000 articles a day.

