About

AI research interests

I'm interested in how machine-learning models learn and reason, how to evaluate them carefully, and how to make inference more efficient. I want my next projects and open-source contributions to involve substantial model work and experiments: implementing methods, comparing baselines, and explaining what the results support.

My completed work so far is grounded in software and quantitative research. I've contributed merged Rust and Python changes to NautilusTrader, a C++ fix to MLX, a fix to four jax.scipy.stats distributions in JAX, forecasting metrics to Darts, and variance-swap bindings and numerical tests to QuantLib-SWIG. Those projects taught me to work in unfamiliar codebases, respond to maintainer review, and make numerical behavior explicit.

Software and quantitative work

trading-algo is my personal research and broker-operations system for US stocks and crypto. Much of the work is validation, data handling, order safeguards, and auditability. The project page documents what the system does and where its evidence stops; I do not publish private backtest returns as proof of performance.

There's an Indian-market sibling, kite-algo, on Zerodha Kite Connect. It provides an operator CLI, idempotent order flow, reconciliation, risk checks, and audit logs designed around applicable SEBI requirements. The strategy layer is intentionally outside the broker-infrastructure repository.

For public-data access, I maintain worldmonitor-mcp, a TypeScript MCP server and CLI around koala73's worldmonitor dashboard. My part is the interface, reliability layer, and direct handlers for providers such as FRED, SEC EDGAR, GDELT, ACLED, and Treasury Fiscal Data.

Shipped, archived

From March 2024 to April 2026 I built Quirio, an AI assistant for Texas residential real-estate agents. It covered listing search, comparative-market analysis, client notes, and document drafts. A small group of agents tried it during early access. I archived it after failing to find a repeatable way to reach users; the frontend and backend are public under Apache 2.0.

Project Saoirse — NSS Grand Prize, 2026

I led the team that designed Project Saoirse, a 50-page proposal for a 10,000-resident orbital settlement at Mars-Sun L2. It won the Grand Prize at the NSS Gerard K. O'Neill Space Settlement Contest in 2026.

I wrote the Politics, Philosophy & Economics chapter: an exergy-denominated currency (Sciubba's Extended Exergy Accounting), a FERC-745-style conservation market, and a tripartite executive (elected representative, appointed expert, sortition citizen) with time-horizon-weighted voting. Full report on nss.org.

Talks

Presented Saoirse at the Saturday Gala Dinner of ISDC 2026 in Washington, DC.

Open source

My eight merged NautilusTrader pull requests include returning Python statistics in name order (#5432), defining how the Rust backtest engine handles an instrument registered twice (#5407), loading every configured IB instrument before startup completes (#4519), benchmark-relative portfolio statistics in the Rust core (#4251), Up/Down Capture ratios (#4354), and correctness fixes in IBKR resolution, return aggregation, and statistic dispatch. In MLX #4748, my first MLX pull request, I fixed vmap of tri_inv, which ignored the triangular flags and returned the full inverse. In JAX #41960, I made the binom, nbinom, geom and bernoulli log-pmfs in jax.scipy.stats return -inf at a non-integer k instead of a finite value. In Darts #3122, I added consistent zero-division behavior across percentage and range-based forecasting metrics. In ib_async, my fix that gives news ticks their originating contract landed on the next branch, and a fix to the v3 migration guide landed in the 3.0 line as my commit.

The open-source page lists every merged pull request and commit with what it changes.

Press

Interviewed on CBC Radio One's Here and Now Toronto in April 2026 about Saoirse. "High school students win national space contest" — 10 minutes, with my teammate Rehan Jafar.

Mock trial

• OBA/OJEN Davis Cup (Peel Mock Trial) — Best Advocate in 2025 and 2026.
• Central Peel placed 3rd in 2025 and 1st in 2026.

Essay

Why I shelved the LLM trader. Notes on removing the Gemini trader from trading-algo's production loop after reading the recent literature on LLMs and market alpha.

Contact

mahimn.patel@mail.utoronto.ca. I usually reply within a few days.

mahimn · about · rev oct 2026