Over the past 6+ years I have designed and deployed production agentic RAG systems, built multi-service Kubernetes platforms, written reinforcement learning agents from scratch, trained generative models, and implemented neural networks in C++. My personal projects span evolutionary algorithms, spatiotemporal deep learning, GPU kernel development, and autonomous systems.
| Project | Status |
|---|---|
| 🤖 Drone RL | Domain randomisation before sim-to-real transfer |
| 👁️ VIT-Qwen | Loading real weights, verifying logits against AutoModelForCausalLM |
| ⚡ CUDA Kernels | Tiled shared memory matmul, working toward cuBLAS-level performance |
| Project | What it is | Write-up | |
|---|---|---|---|
| 🧠 | System Zero | Self-hosted agentic RAG platform on k3s. Hybrid FAISS+BM25 retrieval, ReAct reasoning loop, MCP gateway, Keycloak auth | Blog |
| 🤖 | Drone RL | Autonomous quadcopter learning to hover via SAC/TD3 with PER. Custom PyBullet sim, physical ZMR250 build, no flight controller | Blog |
| 👁️ | VIT-Qwen | Hand-writing Qwen3-4B from scratch and bolting on a ViT encoder. Inspired by DeepSeek's "Thinking with Visual Primitives" | Part 1 · Part 2 |
| ⚡ | CUDA Kernels | Raw CUDA C kernel development on a GTX 1080. Naive matmul to tiled shared memory, benchmarked against cuBLAS | Blog |
| 🌡️ | NEAT Thermal Camera | From-scratch NEAT implementation evolving HSV colour space parameters for thermal object detection through smoke | Blog |
| 🎨 | LBCGAN | Location-conditioned GAN generating scene-aware images to overlay on thermal contours. Prototype, ongoing | Blog |
| 🖼️ | DCGAN on FFHQ | Deep Convolutional GAN trained on 65K face images. Documents what went wrong as much as what worked | Blog |
| 🌊 | ConvLSTM Flood Prediction | Spatiotemporal deep learning for rainfall prediction. Predicts continuous flood severity fields frame by frame | Blog |
| 🔁 | RL Algorithms | Seven algorithms from scratch: VPG, A2C, A3C, GAE, DDPG, TD3, SAC, all with PER where applicable | Blog |
| 📚 | NLP From Scratch | Tokenisation, TF-IDF, sparse one-hot encodings on BookCorpus, skip-gram Word2Vec, positional encoding, attention | Blog |
| 🖥️ | CNN in C++ | Full convolutional neural network in C++ using Eigen and OpenCV. No Python, no framework. Built Aug 2019 to Jun 2020 | Blog |
Languages: Python · C++ · CUDA C · JavaScript
AI/ML: PyTorch · safetensors · FAISS · BM25 · SentenceTransformers · OpenCV · scikit-learn · NumPy
Infrastructure: Kubernetes · k3s · Docker · Helm · Traefik · KEDA · RabbitMQ · Prometheus · Grafana · GitHub Actions
Cloud/Auth: Azure AI Foundry · AzureML · Keycloak · Cloudflare · Tailscale · OVH
Certifications: Microsoft Certified Azure AI Engineer Associate
