Try to break the result.
Compare against strong baselines, separate evaluation data, and test the assumptions. My quantitative work taught me how easily an impressive result can disappear under a better experiment.
University of Toronto student · Builder · Open-source contributor
Mahimn Patel.
I'm interested in how AI models learn, reason, and perform.
I'm a first-year student in Arts & Science at the University of Toronto, St. George, intending to pursue the Applied Mathematics Specialist. My focus is moving toward AI research, model evaluation, and inference, building on merged Rust, Python, and C++ contributions and hands-on work in quantitative research and software.
UofT · first year
Arts & Science · intended Applied Mathematics Specialist
Selected work
Claims link to evidence
01
02
03
04
05
06
NautilusTrader contributions
Eight merged pull requests in a Rust/Python trading engine, covering backtest instrument re-registration, IBKR instrument loading and requests, benchmark-relative statistics, capture ratios, return aggregation, and statistic dispatch and ordering.
Project Saoirse
Grand Prize in the 2026 NSS contest. I led the 11-person team and wrote the political-economy chapter of a 50-page orbital-settlement proposal.
trading-algo
A personal multi-strategy research and execution system for IBKR and crypto. The interesting work is the boring work: validation, safety gates, auditability, and documenting where the backtests can still lie.
kite-algo
Indian-market infrastructure on Zerodha Kite: an operator CLI, idempotent order flow, reconciliation, audit logs, and tests around broker-specific failure modes.
worldmonitor-mcp
An MCP server and CLI for public-data services. The upstream project does the heavy aggregation; my contribution is the interface, reliability layer, and direct-API extensions.
Quirio
An AI assistant for Texas real-estate agents. A small group tried it during early access; I archived it after failing to find a repeatable way to reach users.
How I work
Three defaultsMake dangerous actions boring.
Live orders pass explicit environment, command-line, policy, and human-confirmation gates. Automated actions remain subordinate to deterministic checks.
Write down what failed.
The LLM trader was removed, the crypto book trails buy-and-hold, and Quirio was archived. Negative results are part of the work, not an embarrassment to hide.
Open source
13 mergedMerged work in trading engines, numerical finance, forecasting, MLX, and JAX. Next, I want to take on larger contributions to AI and ML projects, with ownership of implementation and experimental validation.
Merged · most recent
QuantLib #2927AnalyticBlackVasicekEngine now uses the dividend yield and reads the volatility at maturity instead of at time zero, with new testsMerged Oct 2026 ↗
JAX #41960binom, nbinom, geom and bernoulli logpmf in jax.scipy.stats return -inf for non-integer k instead of a finite valueMerged Oct 2026 ↗
MLX #4748vmap of tri_inv now respects the triangular flags instead of silently returning the full inverseMerged Oct 2026 ↗
Nautilus #5432Python statistics come back in name order, so tearsheet rows stay in place between runs of the same backtestMerged Oct 2026 ↗
Nautilus #5407Instrument re-registration in the Rust backtest engine: identical definitions are no-ops, changed ones are rejectedMerged Oct 2026 ↗
QuantLib-SWIG #928Python bindings for variance swaps and replicating / Monte Carlo pricing engines, with numerical testsMerged Sep 2026 ↗
Nautilus #4519Load every configured IB instrument at startup and fail closed on unresolved inputsMerged Jul 2026 ↗
Darts #3122Consistent zero-division handling across percentage and range-based forecasting metricsMerged Jul 2026 ↗
Nautilus #4251Benchmark-relative portfolio statistics in the Rust core with PyO3 bindingsMerged Jun 2026 ↗
ib_async 01da4f0News ticks carry the contract they came from, landed as my commit on the next branchCommit · Apr 2026 ↗
My own repositories
Selected writing
1 publishedWhy I shelved the LLM trader.
I built a Gemini-driven trader, then removed it after the evidence failed to justify the story. Notes on the literature, the implementation, and what would change my mind. →