I build the things that sit between a stopwatch and a stochastic process:
F1 strategy models, matching engines, volatility surfaces, and the maths that holds them up.

The maths, twice · On track · In the market · Under the bonnet · Hire exchange · Team radio

A race and a market pose the same problem: make a decision under uncertainty, against the clock, with someone faster than you trying to make it first. Most of my repos are one of these five ideas, worked out on one side of the pit wall or the other.

An F1 trilogy, built in race order: predict the tyre, read the car, make the call.

Microstructure: where prices actually get made

Volatility and risk: where prices get questioned

The equations these repos are built on. If you read these before the tables, we'll get along.

SVI smile · svi-scanner. Total implied variance at log-moneyness $k$, in Gatheral's raw parameterisation:

Rough Bergomi · rbergomi-engine. Variance driven by a Volterra process whose Hurst exponent $H$ sits near $0.1$, far rougher than Brownian motion:

Hawkes intensity · hawkes-lob. Every trade raises the odds of the next one, and the effect decays. The process stays stable while the branching ratio $\alpha/\beta < 1$:

Marchenko–Pastur · rmt-covariance. With $N$ assets and $T$ observations, $q = N/T$, the eigenvalues of a pure-noise correlation matrix stay inside these bounds. Anything above $\lambda_+$ is worth a look:

The pit-stop inequality · apex-racing-live-strategy. The back-of-envelope version: box on lap $n$ when what the old tyres will lose over the rest of an $N$-lap race is more than the stop costs. It is optimal stopping in a fireproof suit:

More probability than finance: Projects holds experiments in probability calibration, decision theory and betting.

Greenfield repos are the easy part. Getting a change merged into someone else's codebase is the part I'm proudest of: 11 merged so far, mostly in quant libraries and developer tooling.

Quant and research

• ArturSepp/QuantInvestStrats#192: reconstruct instrument P&L across missing quotes, so contributions still sum to NAV
• ArturSepp/QuantInvestStrats#225: run the offline quickstart with sockets patched to raise, so a network call can never pass silently
• ArturSepp/QuantInvestStrats#227: map every printed quickstart output to the Python object that produced it
• bloomberg/causal-ts#83: guard batched ParCorrGPU precision inverses, falling back to OLS instead of inverting an ill-conditioned matrix
• bloomberg/causal-ts#85: pair all 30 example notebooks with Jupytext scripts, so changes review as .py diffs
• bloomberg/causal-ts#86: set up all-contributors, seeded with every human contributor

Compilers, runtimes and tooling

• swiftlang/swift#93473: retry each clone/update in update-checkout, instead of re-fetching all 51 repositories when one flakes
• affaan-m/ECC#3590: accept JSONC comments in opencode.json while checking config membership during ecc install
• morluto/rea#4219: honour package-main selection for trailing-slash directory references
• morluto/rea#4511: keep CommonJS require(`mod`) dependencies visible in the source graph
• tt-a1i/archify#986: add class, tree, timeline and waterfall promoted-type recipes

A prediction market on my own commit count. A GitHub Action quotes a range every midnight IST and settles the trade the next night.

38 sessions · 89% hit rate · best streak 18 · 24 exact / 6 near / 4 beat / 4 miss

📊 2026-10 earnings report: 31 sessions, 87% hit rate, probability +23.7pts. Another month, another set of predictions nobody asked for.

Off the racing line, but still running.

Open to full-time roles in quant research, trading systems, or anywhere a strategy call is really a maths problem. Motorsport included.

• GitHub: @hemv-857 (you're here)
• Email / LinkedIn: in the profile sidebar

Skip the cover letter and give me a concrete problem: a matching-engine latency bug, a wing of the vol surface that won't fit, or a one-stop that should have been a two-stop. I'll have opinions before the call ends.

_{Box, box. · MIT}