Thomas P.

I help businesses build, optimize, and deploy AI systems, from LLM-powered applications and computer vision pipelines to efficient inference on specialized hardware. I'm a Senior AI Engineer with over 10 years of experience in software engineering, deep learning, computer vision, LLM systems, and AI accelerators. My work spans the entire development lifecycle, from research and prototyping to model optimization, system integration, validation, and production deployment. Some of my work includes: Enterprise AI and LLM Applications Developed an Azure-based RAG email responder for a London-based regulatory reporting company, supporting email classification, information retrieval, and response drafting with human review. I also developed an automated financial report validation pipeline. Computer Vision and Deep Learning Independently developed the complete deep learning pipeline for Biotactic's fish detection, tracking, and counting system, improving accuracy by 30% and achieving near-zero downtime in production. At Infosys, I independently developed the LiDAR road segmentation module for an autonomous golf cart, achieving 95% average accuracy at 83 FPS. AI Compilers and Model Optimization At Roviero, I worked on the AI compiler frontend, neural network operators, INT8 quantization, and model optimization. I developed preprocessing and postprocessing pipelines in C for FPGA demonstrations and provided golden reference implementations for hardware validation. At Synaptics, I contributed to NPU compiler development and developed testing and profiling tools that reduced compiler analysis and validation time by 60%. LLM Inference, Local LLMs and AI Accelerators I deploy and benchmark open-weight LLMs using vLLM and SGLang across heterogeneous hardware, including Intel Gaudi, AMD MI325X, NVIDIA H100, H200, RTX 5090, and other server-class and consumer GPUs. My work includes distributed inference, performance benchmarking, and modeling latency, throughput, and resource requirements across different model architectures and accelerator configurations. I've evaluated models from DeepSeek, GLM, Kimi, and other families for reasoning, coding, and practical performance. I've also explored open-source coding-agent harnesses and compared them with commercial tools such as Claude Code for internal development workflows. I've fine-tuned LLaMA using Unsloth and LoRA for an internal knowledge assistant. Generative AI I've developed generative AI tools for text, image, and video generation, including applications supporting the marketing workflows of a global Web3 organization. Research and Innovation I'm a named inventor on US patent applications related to neural network systems and AI compiler development, and a coauthor of peer-reviewed research on LiDAR-based road segmentation for autonomous vehicles. Core technologies: Python, C/C++, PyTorch, TensorFlow, TensorFlow Lite, ONNX, OpenCV, FastAPI, Docker, Azure, AWS, and LangChain. I'm comfortable working independently or with engineering teams, whether it's building a new AI application, optimizing an existing system, or solving complex problems involving model inference and deployment.