Browse through my complete collection of projects. Each one represents a unique challenge and demonstrates different skills and technologies I've mastered.
Side project exploring multi-agent systems applied to developer workflows. LLM-powered Playwright spec generator with auto-debug loop, multi-agent PRD/Figma reconciliation that flags spec conflicts pre-development, and reusable Claude rules + agent harness conventions. Stack: Claude API + custom MCP tools + Playwright + structured tool calls.
Production agentic system built on Claude Code skill architecture + MCP servers (Figma, Google Calendar). Custom agents: weekly-review (data-driven self-management with circuit breakers), wealth-manager (portfolio analysis across MY/SG/US markets), learning-agent (job-market scraping + skill-radar synthesis), profile-optimizer (LLM-driven LinkedIn/resume optimization), coach-planner (real-time decision support and timetable scheduling). Demonstrates production agent design: tool schemas, prompt rules, structured outputs, hot-reloadable skill registry, multi-source data orchestration.

Multi-tenant SaaS digitizing a 200-student tuition center. Role-based portals, Stripe billing, QR attendance, row-level security in Supabase. Saved 5 hrs/week, $500 MRR from 3 pilot centers.
Real-time meeting scheduling SaaS (Calendly + Doodle reimagined). Architecting sync engine for 1000+ concurrent users with Supabase subscriptions. Stripe recurring billing, sub-200ms availability checks via PostgreSQL optimization and Redis caching.
Built a constant-product AMM (x*y=k formula) as a blockchain course assignment. Solidity smart contracts: liquidity pools with LP token minting, slippage protection (max 2% deviation), front-running prevention with deadline checks. Handled integer overflow in Solidity 0.7 (SafeMath before 0.8 built-ins), implemented Babylonian square-root method in assembly for gas efficiency. MetaMask signing, deployed to Rinkeby testnet (~0.02 ETH gas). Combined with current fintech work, demonstrates crypto-fintech bridge for AI agent applications.
Final year capstone: cross-platform mobile app with real-time GPS tracking, offline-first architecture, and Firebase sync. Handled 10MB+ map tiles via LRU cache (100MB limit), reduced battery drain by switching from 1s polling to 10s + geofencing, managed booking state machines across 5 screens. Presented to 50+ industry professionals.
Parallelized image processing (45+ s/image on single thread). Image chunking with Dask + Python threading for I/O. Achieved 8x speedup on 8-core machines, demonstrating Amdahl's Law in practice.