ML Systems & Edge AI Engineer | Building production AI that runs on real hardware, not slides.

I spend most of my time optimizing models until they stop complaining — reducing FP32 bloat into efficient INT8 inference, squeezing detection pipelines on edge accelerators, and architecting memory systems that let AI agents reason without token-bombing the context window.

3 years production. Currently at NXP, Ex-OnePlus. Published ICPR & ACCV . Now more into agentic AI, LLM reasoning, and context management infrastructure.

• ⚙️ Model Optimization — PTQ, QAT, Mixed Precision, deployment-aware quantization
• 🚀 Edge Deployment — NPUs, ONNX, TensorRT, hardware-aware inference pipelines
• • 🧠 LLM & Agentic Infrastructure — Memory systems, context management, reasoning workflows
• 👁️ Computer Vision — Object detection, classification, segmentation (mostly everything)

• 🧠 MemoryClaw — Open-source hierarchical memory layer for AI agents. Four-tier model (recent, important, consolidated, search index) with hybrid keyword + vector retrieval. Cuts token overhead by avoiding brute-force context dumps. Built for OpenClaw framework.

• 🎙️ Maestro — Interactive voice AI tutor generating real-time visual aids (mind maps, diagrams, timelines) during lessons. Warm stone/amber design, horizontal scroll interface. Companion effects library (maestro-effects) for dynamic visual rendering. MemoryClaw started as a personal itch that became useful to others.

• 🤫 More coming soon.

ICPR · ACCV — Computer vision publications. Detection and classification on constrained hardware. The work that came before "everyone" decided AI was easy.

ONNX Runtime · TensorFlow · CUDA optimization · Hardware profiling

_{I optimize models until they stop complaining.}