— A LETTER

I build things
that have to actually run.

Toronto · 2026

I studied Mathematics & Physics at the University of Toronto because I wanted to know what counts as proof. My Master of Science in Civil Engineering at Lakehead University in Thunder Bay is why I work in construction and real estate, with northern builders and developers like ModBox and McRae Development.

What I bring to that world is AI and data analysis that actually ships, aimed at real problems rather than slide decks: a sales console that shows where every lead came from, quotes that can be traced back if they leak, market data measured straight from the order book. Things someone relies on.

A few things I hold to.

─ Measure before you claim

On hyperliquid.build, the 9:00 a.m. premarket call matches the real opening gap 87.8% of the time. Most of that comes from following ordinary US premarket trading, so the page says so, and points to the 4:00 a.m. call (78.6%) as the part only a 24/7 market can offer. An independent re-check found the opening price had been sampled a few seconds too early, which inflated the hit rate by 1.5 points. A number goes on a page only after I have written down what would make it wrong.

─ Backtest before ship

I nearly shipped a foundation-model ensemble for the Champions League. An 83-tie, five-season backtest against a tuned Elo and Poisson baseline showed it added no skill, so Elo went to production and the models went behind a flag. The model then got Arsenal to the final and missed the winner. That is in the write-up too.

─ Keep the load-bearing part small

TaskMarket's settlement layer is one Solidity file with four states and no upgrade path. The arbiter can only send funds to one of the two parties already in the deal. The part everything depends on should be small enough for a stranger to read in one sitting, and small enough for the tests to cover every state.

─ Production, not demo

Tensor Proxies has served paying customers since 2022. ModBox and McRae Development are client sites with real leads behind them, worked in the admin console I built. hyperliquid.build rebuilds and deploys itself every night. None of these has a demo mode. If something breaks, someone notices.

─ Markets are an honest test

I keep coming back to markets, not because I love trading, but because they are the cleanest place to test a forecast end to end. The loss function is honest. You can fool a benchmark; the market does not care what you wanted to be true.

─ Suspicious of three decimals

Physics taught me how much noise a problem can absorb. So on any data project, most of my time goes into finding the honest precision: holdouts that respect time order, calibration, and knowing which digits are real. I would rather show a wide interval than a sharp number I cannot defend.

─ A line under all of it

Find the smallest unit of trust a stranger can check, then build the system around it.

─ Reading lately

Statistical Rethinking · Richard McElreath
Designing Data-Intensive Applications · Martin Kleppmann
Advances in Financial Machine Learning · Marcos López de Prado
Working in Public · Nadia Eghbal