I build high-performance AI systems across the complete hardware–software stack — from GPU kernels and inference runtimes to optimized models, autonomous agents, and verifiable compute infrastructure.

My foundation is in embedded Linux, firmware, real-time systems, hardware design, and silicon bring-up. My current work focuses on Blackwell-native LLM inference, model optimization, agent systems, and confidential computing.

• SparkInfer — native C++/CUDA inference runtime optimized for NVIDIA Blackwell GPUs

• LLM Performance Engineering — NVFP4/FP8 quantization, speculative decoding, long-context inference, KV-cache optimization, tool calling, and RTX 5090 deployment

• Agent Systems — personal AI agents connecting models, tools, skills, memory, communication channels, and device-local actions

• Confidential & Verifiable Computing — trusted execution, GPU attestation, signed compute evidence, fail-closed validation, sealed evaluation, and verifiable sandbox infrastructure

• Physical & Edge AI — Jetson, voice and perception pipelines, embedded Linux, IoT, local automation, and hardware-integrated intelligence

• CUDA/C++ kernel and inference-runtime optimization
• NVIDIA Blackwell architecture and quantized inference
• LLM serving, scheduling, memory management, and latency optimization
• Long-context inference and speculative decoding
• Agent runtimes, tool execution, memory, and multi-step workflows
• Confidential computing, TEE/GPU attestation, and proof-backed execution
• Embedded Linux, firmware, and hardware–software co-design

A complete AI systems stack connecting:

hardware → kernels → runtime → optimized models → agents → verified execution → physical systems

My long-term goal is to design a custom inference accelerator optimized for agent workloads.