The archive of
production systems.
A curated selection of projects spanning artificial intelligence, full-stack web applications, and system architecture. Scroll to explore.
Vigil
Real-time sports coaching system (cricket-first) combining skeleton extraction (MediaPipe Pose + YOLOv8-Pose), multi-person tracking (ByteTrack + DeepSORT), and XGBoost + LSTM biomechanical risk scoring.
- Synthetic data pipeline: annotated training sets from unlabelled match footage; >70% labelling-cost reduction.
- TimescaleDB for biomechanical telemetry; DVC for dataset versioning; Label Studio + Gemma 4 (Ollama) for semantic auto-labelling.
- Privacy-preserving edge inference (ONNX + TensorRT) — raw video never leaves the field, GDPR-friendly deployments.
EdgeShield
Modular framework for deploying quantised (INT8/FP16) computer vision models on edge hardware with fully local inference — zero data transmitted to cloud.
- Benchmarked ONNX Runtime vs TensorRT vs OpenVINO on Raspberry Pi 4 and Jetson Nano across latency, throughput and power.
- Differential privacy noise injection at the feature-extraction layer; validated against GDPR Article 25 (privacy by design).
- Packaged as a Python library with a FastAPI control plane for remote model updates without raw data exposure.
TRIAGE-X
Multi-model triage ensemble (XGBoost + Transformer + LLM) with 0–100 risk scoring, SHAP explainability, <2s latency and 93% AUC-ROC.
- RAG-powered medical report summarisation with hallucination detection; faithfulness score 0.88.
- LSTM + Transformer fusion for predictive vital monitoring with 2–4 hour early warning.
- Microservices architecture (FastAPI gateway + 4 services); PostgreSQL + pgvector for semantic patient-history retrieval.
RETINA-Q
Dual-modality system combining PennyLane 8-qubit quantum circuits with EfficientNet-B0 classical deep learning for OCT and Fundus classification of Central Serous Retinopathy.
- U-Net macular segmentation with BCE + Dice + Tversky loss.
- Grad-CAM + quantum feature importance for explainability.
- PostgreSQL 16 backend; MLflow v2.10 model registry.
Tesserin
Skill-based agent architecture (summariser, linker, critic, retriever) transforming unstructured notes into a queryable Zettelkasten knowledge graph.
- Hybrid storage: ChromaDB (dense) + Pinecone (cross-session) + PostgreSQL (metadata, relationships, versioning).
- LangChain + OpenAI function calling routes NL queries dynamically.
VeriSpec
7-module pipeline: ingestion → OCR extraction (pdfplumber + EasyOCR + NER) → ontology alignment → normalisation (ksi→MPa, F→C) → ASTM/ASME/API/MIL rule engine → append-only audit log.
- PASS/FAIL output with full rule trace.
- PostgreSQL-backed traceability store, Docker Compose deployment, React/Next.js frontend.
SPECTER
Zero-dependency framework installing 18 security skills and 22 enforceable guardrails into any LLM-powered coding agent. Supports 8 platforms including Copilot, Cursor, Windsurf and Cline.
- Skills cover governance, STRIDE/PASTA threat modelling, OWASP Top 10, cryptography, incident response and infrastructure security.
QKD-Lab
Complete BB84 protocol simulation with Eve intercept-resend attack, configurable interception probability, realistic fiber attenuation, Shor-Preskill security bounds and QBER monitoring.
- Monte Carlo analysis, parameter sweeps, CSV export.
- Rust Tauri wrapper; self-contained binary for Windows, macOS and Linux.