AI Engineer who ships. Not "I watched a tutorial" ships. Production apps, real users, actual problems solved.
I build AI-powered products that people use. Led 8 interns to 2nd place at ISSP Expo 2025. Currently shipping at Build Launch Iterate.
π Stolk β Twitter for Stocks
Save 10+ hours/week crawling through news. Get AI-powered sentiment analysis instead.
Real-time community sentiment aggregation with Claude AI analyzing quality posts, extracting themes, and summarizing discussions. News sentiment from 200+ articles across 20+ trusted sources.
The flex: <1s latency | $0.001/analysis | 64% improved sentiment accuracy
Next.js 16 React 19 Claude AI Prisma Neon PostgreSQL Inngest Clerk
π§ X Recommendation Algorithm β Twitter's Ranking Pipeline, Rebuilt
50K tweets. 338K engagements. Two-Tower neural network. Your personalized feed in milliseconds.
TypeScript reimplementation of X/Twitter's recommendation algorithm. 5-stage ranking pipeline (retrieval β hydration β filtering β scoring β selection) with a Two-Tower neural network trained on 676K samples, exported to ONNX for TypeScript inference. 10 filters, 4 scorers, author diversity enforcement, and user-tunable algorithm weights. Trains in Python, runs in the browser.
The flex: 676K training samples | 3.6KB ONNX model | 50Kβ50 tweet ranking | 6 engagement predictions
Next.js 16 React 19 Supabase PyTorch ONNX Runtime Tailwind CSS Vitest
π Self-Improving Prompt β AI Prompt Engineering Toolkit
Define prompts, test them, let AI agents analyze failures and improve automatically.
Multi-agent system that creates a feedback loop: define prompts β evaluate against test cases β AI analyzes failures β generates improved versions β repeat until 90% success rate. Converts user feedback into synthetic test cases to prevent recurring issues. Full-stack with separate FastAPI backend (async SQLAlchemy, Redis workers) and Next.js frontend.
Next.js 16 FastAPI PostgreSQL Redis Google Gemini SQLAlchemy
ποΈ KID-WATCH β AI Study Monitor
Real-time posture, phone, and gaze detection scored into student engagement metrics.
Production proof-of-concept for Viettel's 700,000 home camera deployment. Combines three computer vision models (MediaPipe Pose, MediaPipe Face, YOLOv8) into a weighted engagement score (0-100%) running at 20-30 FPS. Intelligent temporal smoothing, aspect ratio filtering, confidence thresholds, and session reports with personalized recommendations. Single-file Python app, fully offline, zero cloud dependencies.
The flex: 700k camera deployment target | 3 ML models in real-time | 20-30 FPS
Python OpenCV MediaPipe YOLOv8 NumPy
AI-powered deal management system for a Canadian mortgage brokerage. 27 staff, 50+ brokers.
289-commit enterprise platform that syncs CRM data, generates proposal PDFs, and analyzes borrower credit profiles using Google Cloud Vertex AI. Background job processing with Inngest handles 5-13MB PDF extraction pipelines. Multi-zone deployment architecture with real-time data synchronization and enterprise-grade infrastructure (SOC 1/2/3, ISO 27001).
The flex: 289 commits | AI credit analysis | multi-zone deployment | real-time CRM sync
Next.js 16 React 19 Supabase Inngest Google Vertex AI Prisma Claude API
Internal tool β restricted access
β¨ Vibe β AI Website Builder
Describe what you want. Get a full application. Yes, actually.
Multi-agent AI system that generates production-ready React apps from natural language. Sandboxed E2B environments for safe code execution. Ships complete TypeScript + Tailwind applications.
"What, you really believe that? AI? Generating production-ready code?" β Try it yourself.
Next.js 15 tRPC E2B Sandbox Claude 3.5 Inngest Agent Kit Prisma
π¨ Portfolio β Personal Site
3D WebGL effects, buttery smooth scroll, Sanity CMS β built on darkroom.engineering's SatΕ«s framework.
Custom portfolio with Three.js noise wave effects, GSAP + Framer Motion animations, and Lenis smooth scrolling. Content managed through Sanity CMS so I never touch the code to update projects. Built with Bun, strict TypeScript, and Biome for formatting.
Next.js 16 React 19 Three.js GSAP Framer Motion Sanity CMS Tailwind CSS
π¦ Image Style Transfer β WebGPU + Rust + ONNX
Upload an image. Pick an art style. Watch Rust + WebAssembly transform it in your browser.
5 ONNX neural style transfer models (Udnie, Candy, Mosaic, Rain Princess, Pointilism) plus an Adversarial Inception v3 classifier β all running client-side via Rust compiled to WebAssembly. Zero server, zero cloud, zero latency.
The flex: 6 ML models | 100% client-side | Rust β WASM | zero cloud dependency
Next.js 14 React 18 Rust WebAssembly ONNX Tailwind CSS
π§ͺ Autoresearch β Autonomous AI Research Agent
Fork of Karpathy's autoresearch. AI agent modifies a GPT training script, trains for 5 min, keeps or discards, repeats overnight.
Ran Karpathy's autonomous research framework on Apple Silicon (M-series Mac with MPS). The agent ran 22 experiments overnight β modifying architecture, hyperparameters, batch size, and optimizer settings. Best result: 24.5% improvement over baseline (val_bpb 1.924 β 1.453). Key findings: halving batch size from 65Kβ16K gave the biggest single win (β7.0%), optimal depth was 3 layers at 10.7M params, and learning rate tuning (MATRIX_LR 0.06) squeezed out the final gains.
The flex: 22 experiments overnight | 24.5% val_bpb improvement | 10.7M param model | Apple Silicon MPS
Python PyTorch MPS (Metal) Claude Agents
- π Call Analysis β AI call intelligence platform I built with 8 BCIT interns (2nd place ISSP Expo 2025)
- π’ PortPal β Mobile-first PWA for ILWU longshoremen to log shifts, track earnings, and hit income goals (private β client project)
- ποΈ AI Customer Support Agent β Voice-enabled RAG with Pinecone, Firecrawl, and Vapi.ai (demo)
- π LLM Router β Route prompts to the cheapest model that can handle them (demo)
- π΅ Viet Bike Scout β Parallel browser agents scrape 18 motorbike rental shops across Vietnam simultaneously
- π₯ Clinic Book β Scrape physio/massage/chiro availability across Vancouver with SSE streaming
- π Velocity β Real estate deal management with multi-LLM analytics (Claude, Bedrock, Gemini) (private β client project)
- β¨οΈ Kolm Dashboard β Business intelligence for Kolm Kontrast sauna studio (private β client project)
- π¦ Blue Pearl β Mortgage marketing site with debt calculator, Google Maps, and SEO
- π Whiteboard β Shift tracker for Douglas College IT dept β 4 years of production use (PHP/MySQL)
- π Build Launch Iterate β My dev studio's marketing site
Led 8 BCIT interns β 2nd place at ISSP Expo 2025
Built an AI call analysis app from scratch. The wild part? None of them knew Next.js or TypeScript 3 months prior. They normalized a nightmare dataset (200 rows Γ 220k columns of 10-year manual entries) into clean, queryable data. Result: <10ms queries, modern dashboard, AI-powered call insights for sales teams.
Languages: TypeScript β’ JavaScript β’ Python β’ Rust β’ SQL β’ C++ β’ PHP
Frontend: Next.js β’ React β’ Tailwind CSS β’ shadcn/ui β’ Three.js β’ GSAP β’ Framer Motion
Backend: FastAPI β’ Prisma β’ tRPC β’ Inngest β’ Node.js β’ Sanity CMS
AI/ML: Claude API β’ Google Vertex AI β’ Gemini β’ OpenAI β’ AWS Bedrock β’ PyTorch β’ ONNX Runtime β’ Pinecone β’ MediaPipe β’ YOLOv8 β’ OpenCV β’ RAG
Infrastructure: Vercel β’ Supabase β’ Neon PostgreSQL β’ Redis β’ Stripe β’ E2B β’ WebAssembly
Building something interesting? I'm always down to chat.
I don't do LeetCode. I build products.