Selected Works
Production systems, open-source libraries, hackathon winners, and peer-reviewed research. Filter by category below.
AuraHealth
arrow_outwardProduction voice-first triage platform that pre-authorises emergency healthcare payments into escrow before the patient arrives. Node.js/TypeScript APIs, Interswitch payment rails with OAuth 2.0 + HMAC-SHA-512 webhooks, three real-time SSE dashboards over one PostgreSQL source of truth.
Voxtar
arrow_outwardHealthcare voice AI platform. Deploy domain-specialised voice agents for triage, patient follow-up, and post-op monitoring. Multi-service: Next.js frontend, Python/FastAPI backend, self-hosted LiveKit for real-time audio, Qwen3 TTS and ElevenLabs for clinical-grade synthesis.
Vectorless
arrow_outwardDocument retrieval for the reasoning era. Structure-preserving retrieval that lets LLMs reason over document maps instead of vector search. No chunking, no top-K, no vector DB. Live at vectorless.store.
Vectorless Engine
arrow_outwardThe Go retrieval engine that powers Vectorless. Reasons over document structure — not embeddings. No chunking, no top-K, no vector DB. The opinionated take that kicked off the project.
Hypatia
arrow_outwardFull SaaS AI learning platform. NestJS + DynamoDB + Next.js. Tiered subscriptions ($20 / $50 / mo), credit-based AI usage metering, personalised learning-path generation via LLM orchestration.
Coursified
arrow_outwardEnd-to-end RAG pipeline transforming YouTube content into structured interactive courses. Django backend, Next.js frontend, full SaaS with authentication, custom reporting, and AI content structuring.
Context8
arrow_outwardCollective problem-solving memory for coding agents, powered by Actian VectorAI DB. Context7 gives your agent the docs. Context8 gives it what the docs don't cover — tacit knowledge, gotchas, working configurations.
Aurasense
arrow_outwardReal-time agentic voice AI. Time-aware Graphiti RAG memory for long-running agentic interactions. Multi-model orchestration (Llama + Grok) with sub-second Groq TTS latency. FastAPI backend, Next.js frontend, Neo4j knowledge graph.
Stealth AI Platform (Contract)
Production multi-tenant agentic SaaS for US businesses — dedicated phone numbers, isolated agent configurations, self-service dashboards. LangGraph + LangSmith tracing + Opik versioning. 200+ synthetic eval datasets before production deployment.
Oncolens
Edge-native pathology AI. Applied Knowledge Distillation to compress PathFoundation and UNI foundation models into a sub-300KB MobileNetV3 — 96% sensitivity at 3.6ms inference. Human-in-the-loop escalation for low-confidence diagnoses.
llmgate
arrow_outwardLiteLLM for Go. Provider-agnostic LLM client over Anthropic, OpenAI, Gemini with router, fallback, cost tracking, capability flags, and composable middleware. Published on pkg.go.dev.
swarmtorch
arrow_outwardPeer-reviewed PyTorch-native library implementing Particle Swarm Optimization, Grey Wolf Optimizer, and related metaheuristic algorithms. Rigorous mathematical peer review, clean API, active maintenance.
tytube
arrow_outwardTypeScript port of pytube. Extract YouTube metadata, list streams, and read captions without the YouTube Data API. Works in Node, Bun, and the browser. Published to npm.
youtube-transcript-ts
arrow_outwardFetch YouTube transcripts and subtitles from Node.js, Bun, and Deno. TypeScript port of youtube-transcript-api. Zero external dependencies.
pyzheimer
Python library for speech-based neurodegenerative disease detection. Wraps Google's HeAR health acoustic foundation model plus classical acoustic features — reusable across RLN monitoring, Parkinson's, Alzheimer's, and respiratory indications.
Scrunchy
arrow_outwardDecentralised game assets on Solana. NestJS backend for asset creation and transaction management, Rust Anchor smart contracts, Python Typer CLI for developer integration. Full four-repo ecosystem.
HealthDAG
arrow_outwardDecentralised health data platform on blockchain. Architected and deployed end-to-end in a 48-hour hackathon sprint — DAG-based consensus for health records, four-repo distributed architecture.
ChainVote — Privacy-Preserving Voting
arrow_outwardPrivacy-preserving blockchain voting system in Rust on the Partisia blockchain. Zero-knowledge voting primitives. Most Valuable Participant at Partisia Blockchain Hackathon 2024.
BridgeHook
arrow_outwardWebhook testing in your browser. Cloudflare Workers + Neon PostgreSQL — inspect, replay, and debug webhook payloads without local tunneling tools. Designed for the 'just paste the URL and start receiving' workflow.
Stroke Lesion Segmentation
Training a medical image segmentation model on AWS SageMaker to identify ischemic stroke lesions on CT scans. Grant-funded work with Duke Global Health Institute — distributed cloud-scale ML on real clinical imaging datasets.
Biometrics Proctoring Backend
Distributed microservices backend built from scratch: face recognition engine, liveness detection, exam session manager, real-time event bus — all communicating via gRPC with fault-tolerant inter-service contracts. Full Terraform IaC for GCP; pgvector for face-embedding similarity search at scale.
CancerAg — GPCR Biased Agonists
Structure-aware machine learning pipeline for identifying biased agonists at G-protein-coupled receptors — therapeutic potential in cancer treatment. First-author publication at GIW/ISCB-Asia 2025, Hong Kong, and Briefings in Bioinformatics (OUP).
HIV-1 Drug Resistance Genomics
Comprehensive analysis of 741,873 HIV-1 sequences from 47 African countries — mutation trends, subtype diversity, and therapeutic implications (1990–2024). First-author publication in Briefings in Bioinformatics (OUP), December 2025.
Breast Cancer Histopathology Classification
arrow_outwardML pipeline using Google's Pathology Foundation Model for H&E-stained microscopy analysis. 89% accuracy distinguishing malignant from benign breast tissue. Winner of Hult Prize 2025 at UCH.