Building AI systems across consumer-GPU LLM training, GPU-accelerated computing, autonomous agents, and edge ML.
|
Consumer-GPU LLM training & evaluation Adapting nanochat to push end-to-end LLM training on a single RTX 4070 12GB — including ~1B-parameter training, OOM-safe fallbacks, checkpoint recovery, LoRA, lower-memory evaluation, and experiment automation.
|
NVIDIA Hackathon Winner · GPU graph intelligence A disruption early-warning system for London small businesses. Uses RAPIDS cuGraph/cuDF to propagate live transport and road disruptions through urban networks and identify affected businesses and critical infrastructure.
|
🏆 StellaVercel Hackathon Winner · Autonomous agent system A multi-agent system that runs real small-business workflows end-to-end, combining autonomous planning with deterministic financial logic, compliance gates, memory, human approvals, and a full operational console.
|
Final-year research · Wireless sensing + Edge ML Human activity and posture recognition using WiFi Channel State Information, combining signal processing, deep learning, and ESP32-based wireless sensing.
|
- ⚡ Efficient LLM training & inference on constrained hardware
- 🧩 CUDA / Triton / GPU systems
- 🤖 Reliable agent orchestration and autonomous systems
- ☁️ ML infrastructure, serving, and distributed systems
- 🔬 Small, focused experiments for understanding modern model architectures