Oscar Daniel Navarro Cieza

Trading strategies reimplemented in Python and put through continuous paper trading to find out which ones survive contact with the live market.

The system exists to rule strategies out, not to make money: it measures each one’s live expectancy against its backtest and flags the ones that stop reproducing what they promised. Most failing is the expected result — a strategy that holds up outside its own backtest is the exception, and finding one means measuring the rest honestly. Over 40 strategies run against Binance perpetual futures, with a REST API exposing equity curves and the live-versus-backtest evaluation. A pre-push hook blocks any commit where the live signal path stops reproducing the backtest. Paper trading only — it has never traded real money, and its own documentation warns against reading the backtest Sharpe as a live expectation.

Photograph handwritten class notes, get a study summary and an auto-graded exam.

A three-stage pipeline takes the photographs through transcription, summary and question generation. It uses context-window-aware chunking rather than retrieval, deliberately: the product needs full-document coverage, not a relevant slice. Multi-user with authentication, installable as a progressive web app and as an Android app. 168 tests, and a security audit whose ten findings were all fixed with regression tests covering each one.

A weight tracker with no server and no account.

SQLite compiled to WebAssembly and persisted in the browser, with charts drawn by hand in SVG and no charting library. One codebase ships as a progressive web app and as an Android app. Its own SECURITY.md states the two honest limitations: the database is not encrypted on the device, and the published APK is signed with a debug key rather than a release key.

A crypto trading system built to answer one question — is there an edge after real costs? — and closed when the answer was no.

Its go/no-go criteria went into the first commit, before any signal or backtest code existed. The original range-reversion thesis failed them: over three years and 12 coins, net of costs, every configuration lost money, and its conviction score was inversely calibrated. Five directional strategies that had passed walk-forward and double out-of-sample checks were then run live in dry-run: 110 trades over 7.8 weeks returned -64.2R, and of 22 combinations positive out of sample, 2 were positive live. Closed on 3 August 2026 at v1.0, with no real capital ever traded. Its own conclusion is that the strategies have no edge. The signal code is shared by the backtest and the live monitor, so what was validated is what ran.