Engineer → SaaS founder/operator → MBA → investment banking. Somewhere in there I picked up the habit of not trusting a number until I've rebuilt the model myself — which, if I'm honest, is really just an excuse to keep writing code for a living after finance told me to stop.
Based in India. On a given night I am equally likely to be arguing with an EBITDA bridge in a spreadsheet or with a segfault in a terminal, and neither argument is going noticeably better than the other.
A DCF and a git rebase have more in common than people think: both are
just you insisting, with great confidence, that the past can be tidied up
into a story that leads cleanly to the number you already believe.
The question I keep asking: does the tool actually help me understand the business, or does it just make the output look confident? Most finance software optimizes for the second one. I'd rather build the first — even if that means the "financial model" is a Python script with opinions.
|
Face detection (YuNet) + embeddings (SFace), clustered with a greedy online pass and then a proper average-linkage recluster — and once you manually correct a face, it's pinned: the algorithm is no longer allowed to have opinions about that one.
|
LBO modeling, DCF analysis, deal-analysis tooling — the thing I actually underwrite with, not a portfolio-piece demo. If a sensitivity table lies to you, it's this one's fault, and I'd want to know.
|
Cycle-accurate FPGA datapath sim: deterministic callback scheduling, PCS encode/decode, a five-level order book, zero heap allocation on the hot path — because "close enough" isn't a real answer at the clock-cycle level.
|
|
Named after the real second-order Greek. Prices options from first principles (Black-Scholes, a binomial tree for early exercise) instead of bucketing historical fills, then decomposes every trade's P&L into delta/gamma/theta/vega/vanna/volga via a Taylor expansion — with the leftover reported honestly as residual, not hidden in whichever bucket makes the total look clean.
|
The mid-price lies to you the instant the book is imbalanced; the micro-price is what you get when you stop pretending the best bid and best ask are equally likely to trade next. Implemented as the actual Markov chain over the queue-imbalance state space — not a smoothed heuristic — because an estimator you'll trade on should survive being asked why it said what it said.
|
Grew out of a fix for LiuAlgoTrader (abandoned since 2023, still north of 900 stars) that wanted a Postgres server just to backtest a moving average. This needs a venv and nothing else. The backtest engine and the live-trading path share the exact same
|
QuantLib — the C++ library half of quant finance is quietly built on — had a NaN hiding in its Gauss-Laguerre quadrature: past order ~200, one of the weights underflows to exactly 0.0, and inf × 0 in IEEE 754 is NaN, no questions asked. The fix isn't "add an epsilon and pray," it's re-deriving the weight in log-space so the underflow never has anywhere to hide — tested against the real consumer (AnalyticHestonEngine) and independently re-derived in NumPy just to be sure I wasn't fooling myself.
edgartools — much smaller, and I'll say so: the quickstart claimed Python 3.8 while pyproject.toml actually required 3.10, and its own docs quietly recommended a cash_flow_statement() alias that's deprecated for removal in v6.0. No math, just paying enough attention to notice the docs were lying to new users — merged.
One of these is a rigor problem, the other is a reading-comprehension problem. Both count.
ever-gauzy — a null-prototype object — the exact defensive pattern you'd reach for specifically to dodge prototype-pollution bugs — crashed the utility function written to guard against prototype pollution. Object.create(null) has no .constructor to read .name off of; the fix is one getPrototypeOf check. Merged.
ever-gauzy — their own security-disclosure badge pointed at a domain that no longer resolves (DNS SERVFAIL, not a 404 — the kind of dead link most link-checkers miss). The irony of a broken link on the "here's how to report a security issue" line felt worth fixing quickly. Merged.
Riskfolio-Lib — the constructor built its own returns setter — type check and all — then routed around it entirely: self._returns = returns, no validation, ever. A single NaN anywhere in your input data doesn't fail at construction; it fails four stack frames deep inside scipy.linalg.eigh, with a message that has nothing to do with your actual data.
Riskfolio-Lib — Axes.plot_date and matplotlib.cm.get_cmap were both quietly removed in Matplotlib 3.11. requirements.txt still says >=3.9.2. Anyone on a current install can't plot a portfolio without hitting an AttributeError first.
trade-frame — 37 images in the README that were never images —  pointing at .h files and directories, rendering as broken-icon confetti through the project's own pitch. Two of the targets had also been renamed .h → .hpp years earlier and nobody had updated the doc since.
trade-frame — asked where the meaning of an IQFeed trade-condition code comes from. Answer: nowhere in the repo, on purpose — IQFeed publishes the code table live over the same socket connection, so hardcoding it would just go stale. Documented the actual mechanism instead of inventing a table that would be wrong by next quarter.
matching-engine — an out-of-range price and a genuinely malformed one got rejected with the identical error reason — fine, until you're debugging real order flow at 2am and the log gives you no way to tell which one actually happened.
py_lets_be_rational — no LICENSE file, which is a small thing right up until it silently blocks conda-forge packaging for everyone downstream.
cody-special — erf_cody and normaldistribution got split into their own files at some point, and the @numba.njit decoration that made them fast didn't make the move with them.
jev-ultrafast — the agent's own browser target opens on about:blank, which is already readyState == 'complete' before the real navigation even starts — so the very first poll after Page.navigate could read that stale state and hand the policy a page that never loaded. It terminates the run as BLOCKED in under a second, which looks exactly like the agent failing when the harness never actually observed anything real.
awesome-systematic-trading — checked all ~650 links in the list against the GitHub API and actual DNS resolution, not just an HTTP status code (which false-positives constantly on ordinary bot-blocking). Six repos were genuinely gone; no guessed replacements went in for the ones without a verified successor.
wilsonfreitas/awesome-quant — Add vanna to Financial Instruments & Pricing (open).
NandhaKishorM/laya — Fix single-option choice question crash in DecisionModel.forward (merged).
PyPortfolio/PyPortfolioOpt — Add return_raw option to BlackLittermanModel.bl_weights (open).
wangzhe3224/awesome-systematic-trading — Add vanna to Pricing (open).
