Final-year Computer Science (AI) at McGill, graduating May 2027. I joined a royalty company listed on the TSX and NYSE as its first in-house developer and took its AI and data systems from an empty repository into production over one summer — which in practice meant learning to read mining technical reports.
I'm most useful somewhere the engineering touches a domain with real rules in it: finance, compliance, anything where being wrong has consequences and the right answer has to be traceable.
Portfolio Ten projects, each with the problem, the architecture, and what I'd change. Next.js 16, 19 static routes. |
TaxMapCA Four questions, one exact Canadian sales-tax rate, a government source behind every answer. No model in the decision path — a tax answer that can hallucinate is worse than none. |
TabScribe Research workspace running entirely on-device through Chrome's Gemini Nano. Summarises, rewrites, translates and cites with the network off. |
CodeRoast AI code review delivered through the editor's own diagnostics pipeline. Linters catch syntax and miss judgment; tools that catch judgment write reports nobody opens. |
The install counts and ratings above are live endpoints — they read from the stores, not from me.
VectralQ — enterprise search on Postgres
instead of a vector database. Three retrieval lanes — weighted BM25, pgvector, and a phrase lane
scored separately from bag-of-words — fused with Reciprocal Rank Fusion over ranks rather
than scores, then reranked by a cross-encoder. Tenant isolation enforced by Postgres row-level
security with FORCE: a query that forgets the tenant returns nothing instead of returning
someone else's documents.
Into McGill TechAccel within a month, piloted, shut down on weak demand. The retrieval worked; the market didn't. Killing it was the right call.
ARMM — a market maker that decides what kind of
market it's in before it decides how to quote. BAND/DRIFT/EVENT classification from rolling
quantiles, EMA persistence and robust volatility, with hysteresis and an event cooldown so it
can't whipsaw itself. Inventory is a steered control variable, slew-rate-limited so the strategy
can't chase its own signal. Solo in 24 hours — 2nd place, National Bank Challenge.
wtf — diagnoses terminal errors against your actual
machine: your Python, your installed packages, your requirements.txt, whether your virtualenv
is even active. Pasting a traceback into a chatbot throws all of that away. Runs fully offline
with no API key.
DocuMint — README generation that survives a real repository. The chunker and context-aware reader are the project; everything else is plumbing around deciding what to put in front of the model.
I also built a Chrome extension that files a public GitHub issue every time I get distracted. It worked, which was the worst part.





