Software developer working across quant finance, real-time systems, and AI tooling.
I write code; occasionally it works.
Most of my repos are experiments. These are the ones I'd defend in a review.
| Project | What it is | Why it's here |
|---|---|---|
| limit-order-book | Deterministic limit order book & matching engine in C++20 | No locks on the hot path; determinism is tested, not hoped for |
| svi-scanner | SVI volatility-surface fitter that scans Deribit options for arbitrage | Fits a surface, then looks for where it disagrees with the market |
| cognitive-convergence-detector | Detects when institutional signals converge — crowded positioning / systemic risk | Correlation as an early-warning signal, not dashboard decoration |
| mcp-vetter-agent | AI agent that audits MCP servers before you connect to them | Because I hand agents real tools, so I check what's on the other end |
| edge-mem | Offline-first edge knowledge engine, Qdrant-backed cloud federation | Works with the network off, and still finds things |
| Codeinsight | Ask questions about any repo in plain English; see the architecture it infers | RAG + knowledge graph over code, with the graph drawn out |
The rest of the portfolio — 26 repos, grouped by what they're actually for
Quant & market microstructure
| Repo | What it does |
|---|---|
| lob-engine | Price-time priority LOB in Rust, std-only — plus TWAP/VWAP/POV execution algorithms |
| hawkes-lob | Hawkes process model — tests whether BTC trades self-excite like order books do |
| rmt-covariance | Covariance cleaning via Random Matrix Theory vs Ledoit-Wolf, with walk-forward backtests |
| rbergomi-engine | Rough-volatility pricing — Monte Carlo with control variates, calibrated to live SPX options |
| 151-trading-strats | 151 strategies across 18 asset classes; based on Kakushadze & Serur (2018) |
| market-pipeline | Binance WebSocket → Redis → FastAPI, 35 ms end to end |
AI tooling & agents
| Repo | What it does |
|---|---|
| intellex | Natural-language requests → clean, source-backed datasets |
| vizql | Visual SQL explainer — watch a query execute step by step |
| vizarch | System architecture diagrams from plain English, as SVG |
| commit-archaeology | Commit history → feature eras, who broke what, and the bus factor |
| dependency-time-machine | Any repo + any past date → that project's snapshot, plus what shipped since |
| impact-lens | Impact & sustainability media analysis, before/after comparison |
| J.A.R.V.I.S | Finds issues matching your skills, plans the contribution, drafts the PR |
F1 — a three-part series
| Repo | What it does |
|---|---|
| predictive-tire-engine | Tire degradation predictor, D3 dashboard, live telemetry |
| f1-performance-intelligence | Telemetry dashboard with an AI race engineer |
| apex-racing-live-strategy | Race strategy simulator — tire/fuel modelling, pit optimisation |
Fintech
| Repo | What it does |
|---|---|
| ptm-recovery-agent | Recovers failed Paytm payments, measuring true incremental lift |
| rzp-recovery-agent | Same for Razorpay — causal, not just "recovered count" |
Platforms, native & utilities
| Repo | What it does |
|---|---|
| ORION | Geopolitical risk dashboard — energy supply chains and conflicts, real time |
| textdistance-rs | Rust port of textdistance — 43 algorithms, 77–1788× faster than Python |
| agent-office-app | macOS SwiftUI app running multiple AI agents — streaming, cost, task queues |
| PRESENT | macOS menubar attendance tracker — safe-skip maths and lecture reminders |
| FinCortex_Finvigators | AI CFO assistant — spend, cashflow forecasting, planning |
| Projects | Experiments in probability calibration, decision theory, betting |
| internship-tracker | Scrapes 270+ companies, daily digests via Telegram |
Greenfield repos are the easy part. Pushing on someone else's repo is the part I'm proudest of:
- jaegertracing/jaeger#9820 — separate ACP prompt blocks with a delimiter in the AI sidecar
- topoteretes/cognee-community#202 — Raindrop.io data-source connector
- jankrepl/deepdow#152 —
__sub__/__rsub__on theLossclass
24 sessions · 88% hit rate · best streak 14 · 12 exact / 5 near / 4 beat / 3 miss
📊 2026-09 earnings report: 1 session, 0% hit rate, probability -0.2pts. The alpha is imaginary but the vibes are quarterly-confirmed.
| Date (IST) | Forecast | Shipped | Result | Δ pts | Hire probability |
|---|---|---|---|---|---|
| 2026-10-24 (today) | 3–8 | ⏳ | Market open | ||
| 2026-10-23 | 4–8 | 3 | 🟢 Near hit | +0.88 | 17.58% |
| 2026-10-22 | 3–8 | 5 | ✅ Exact hit | +1.12 | 16.70% |
| 2026-10-21 | 3–8 | 5 | ✅ Exact hit | +1.12 | 15.58% |
| 2026-10-20 | 3–9 | 8 | ✅ Exact hit | +1.39 | 14.46% |
| 2026-10-19 | 3–65 | 8 | ✅ Exact hit | +0.62 | 13.07% |
| 2026-10-18 | 3–62 | 9 | ✅ Exact hit | +0.65 | 12.45% |
| 2026-10-17 | 3–62 | 2 | 🟢 Near hit | +0.53 | 11.80% |
| 2026-10-16 | 3–62 | 4 | ✅ Exact hit | +0.56 | 11.27% |
| 2026-10-15 | 4–62 | 3 | 🟢 Near hit | +0.55 | 10.71% |
| 2026-10-14 | 4–62 | 3 | 🟢 Near hit | +0.45 | 10.16% |
| 2026-10-13 | 4–11 | 62 | 🚀 Beat | +1.80 | 9.71% |
| 2026-10-12 | 4–11 | 175 | 🚀 Beat | +1.70 | 7.91% |
| 2026-10-11 | 4–15 | 8 | ✅ Exact hit | +0.63 | 6.21% |
| 2026-10-10 | 4–28 | 4 | ✅ Exact hit | +0.14 | 5.58% |
How does this work? (a.k.a. why is this in my README)
- Every day at 00:05 IST a GitHub Action forecasts a range for the day: the 7 day trailing window, trimmed — the smallest and largest are dropped, so one monster day can’t turn every forecast into a shrug. It lands about half the time.
- Counts are the commits that still exist on my repos' default branches, so
git logagrees with me. GitHub's contribution graph never retracts a commit — rebased and amended work stays counted forever, and it claimed 74 for a day with 38 commits in it. - The next midnight it fetches my real commit count and settles the trade.
- Points follow the work: 0 for a day with no commits, up to 1 for matching the top of the range, and up to 1.5 for beating it. Streaks add a small bonus.
- Landing inside the range is a hit, 1 under is a near miss, further under is -0.25, and no commits on a day that expected work is -0.5.
- Shipping more is never punished. Beating the top of the range is the best possible day, and the score keeps climbing with the overshoot.
- Market manipulation means one thing only: 10+ commits within 60 minutes in this profile repo — the only way to game the contribution graph. Overachieving a low forecast across real repos is not a crime. Penalty: -1.5 pts.
- Visitors: 👍/👎 on today's voting issue to predict whether the bot hits. No prize, just bragging rights.
- Monthly: on the 1st, an earnings report summarises the previous month.
- Probability is 100% fiction. The 100% real part: I actually ship code, and I'm actually looking for a job.
Languages
Frameworks & Tools
AI & Data
Domains
Open to full-time roles — quant, platform, or AI-infrastructure engineering.
- GitHub: @hemv-857 (you're here)
- Email / LinkedIn: add them to the profile sidebar
Prefer a concrete conversation to a cover letter. Point me at a matching-engine latency bug or a CI pipeline that takes 40 minutes and I'll have opinions before the call ends.
LICENSE — see the file for the full text.
MIT — applies to the profile README and anything in this repo.