I like solving problems, in code and in how models think.

I lead a team of three at EW Analytics, building across GIS portals, management systems and automation tooling in GeoNode, Google Apps Script, React and Java. One of the bigger wins was automation software that cut deployment time by 40% by eliminating repetitive manual workflows. I've also customized the MapStore React UI and built repeatable PostGIS/GeoServer data pipelines that cut manual publishing time in half.

• GeoNode
• React
• Java
• PostGIS
• GeoServer
• MapStore

My focus here is on sparse autoencoders, specifically absorption and splitting.

• Sparse autoencoders
• Mechanistic interpretability

At Algoverse I've been investigating emergent misalignment in language models, from 0.5B up to 32B parameters, using LoRA fine-tuning and activation steering to show it holds up reliably across model scales. I built linear probes on adapter weights that work as early-warning indicators, catching misaligned behavior before it shows up in a model's output. The work has been co-authored and open-sourced on HuggingFace.

• LoRA
• Activation steering
• Linear probes
• HuggingFace

• MechInterp (Spotlight)
• CogInterp
• UniReps (Honorable Mention)

HomeBatch is a marketplace for Edmonton home cooks, built so someone can sell anything from a single dinner to a full freezer batch to a standing weekly subscription. It's still pre-launch, but it's already a real system underneath: 15 feature modules spanning customer, vendor and admin roles, an 8-migration Postgres schema with row-level security, and six Supabase edge functions handling Stripe checkout, payouts and email.

• React
• TypeScript
• Supabase
• Stripe Connect
• Vitest
• Playwright

Study Buddy AI is a bot built to make revising notes feel less like re-reading and more like asking a question. It uses Retrieval Augmented Generation to pull the right passage out of a student's own material, and LangChain to turn that into a natural answer rather than a raw document dump.

• Python
• LangChain
• Google AI
• NumPy

Ace Budget AI came out of a hackathon built around one annoyance: nobody wants to type in every receipt just to see where their money went. It reads receipts and parses emails automatically, then runs a TensorFlow and spaCy pipeline to predict where spending is headed, on a Next.js frontend over a Python Flask backend, containerized with Docker.

• Next.js
• TypeScript
• Python Flask
• TensorFlow
• spaCy
• Hugging Face
• Docker

HandMouseApp started from a simple question: could a webcam replace a mouse? It tracks hand landmarks in real time with OpenCV and Mediapipe, and turns that tracking into cursor movement, clicks, double-clicks and drags through a PyQt5 interface.

• Python
• OpenCV
• Mediapipe
• PyQt5