Projects
Synthetic Focus Group as a Service for Canada
Describe your target audience, paste your content. Get instant reactions from AI personas grounded in 30+ real Canadian data sources. A LangGraph pipeline searches a persistent database of 5,000+ census-grounded personas, generates new ones to fill gaps, and runs concurrent reactions via five LLMs through OpenRouter. Aggregates sentiment, optimizes your message, and re-tests it.
Verifiable Agentic Workflows
A framework that treats AI like an apprentice. Upload a document and the system extracts structured data, shows an ingestion plan in plain language, and waits for human sign-off before executing. Three stages, three models, human checkpoints at every step. The system earns autonomy through measured performance: from hands-on review, to auto-approving high-confidence results, to overnight batch processing with exception-only alerts. First domain: HR attendance compliance with policy extraction, points-based discipline, and full audit trail.
Personal AI Assistant That Learns From Research and Evolves
After completing a 10-day OpenClaw course, I kept building. SAI is what came next: a personal AI assistant I use daily - it sends AI news digests via Telegram, reads my Google Calendar, triages email, and summarizes research papers on a schedule. On top of that, it improves itself: SAI ingests papers into a ChromaDB vector database, uses RAG to propose changes to its own operating rules, and waits for per-diff human approval before applying anything. Built with a 5-layer security model, multi-source research fetcher, and tiered knowledge base. Runs on a self-managed Hetzner VPS alongside monitoring, notes, and status dashboards.
10-Phase Methodology for Applied AI Systems
A generalized framework distilled from shipping MaplePulse, Apprentice, and SAI.
AI & ML Experiments
-
3D Word Embedding Visualizer
Neural network that analyzes IMDb movie reviews and visualizes 64-dimensional word embeddings in 3D space, revealing how similar words cluster together.Watch demo
-
Satellite Image Classification
Transfer learning with MobileNetV2 to classify 21 land use categories from 128x128 satellite images with real-time inference and confidence scoring.Watch demo
-
Real-Time Sign Language Recognition
LSTM neural network trained on MediaPipe keypoint sequences to recognize sign language gestures in real-time from webcam input, with live probability visualization.Watch demo
-
AI Plays Super Mario Bros
Reinforcement learning agent trained with PPO to play Super Mario Bros, learning optimal strategies through millions of gameplay iterations.Watch demo