Six projects, built May–August 2026. Each entry states what the project does and what it actually measured — including the three where the central idea did not hold up.
QuantumSafe Scan · quantumsafescan.com — Finds
cryptography a quantum computer would break, scores exposure 0–100, names the
NIST-standard replacement. The hard part isn't matching RSA; it's not matching it
in a comment or a dead code path. A hybrid AST-and-regex engine with reachability
ranking took precision from 49.1% to 100% on a 9-language benchmark at 100% recall
over 50 seeded cases. Audited 11,275 files; of 37 PyPI packages, 86% had a finding.
Exports SARIF and CycloneDX CBOM, so it drops into real CI.
RL-Trader · 4-page paper — PPO from scratch, aimed at markets as a hostile testbed, asking whether an agent that looks brilliant once survives ten runs. It doesn't. One seed showed +275% against a +19% benchmark; across ten, individual returns span −29% to +226% and the mean, +57% [+16%, +106%], is indistinguishable from buy-and-hold's +34% (p = 0.79) — and significantly worse on equities (p = 0.002). Retrained over four slices of Bitcoin's history, it lost money in all four. Its own significance tooling caught the false positive a naive version ships, reproducing Henderson et al. (2018) in a new domain. Adds an original surrogate-data test separating "the agent is weak" from "there is no signal."
AlphaZero-style Chess Engine · rlchess.xyz — Rules only: no openings, no human games, no evaluation function. Dual-headed ResNet, PUCT tree search, 4,672-move policy head, playable in-browser via ONNX Runtime Web. Scaling self-play 10× bought 14 Elo. Diagnosed as a self-play signal problem, not a compute ceiling — 92% of games ended non-decisive, starving the value head of gradient. A classical alpha-beta control beat the random baseline 10–0, confirming the harness was sound.
AlphaGlyph · alphaglyph.org — Symbolic-regression alpha discovery, evaluated against its own overfitting. Combinatorial Purged Cross-Validation puts the probability of backtest overfitting at 72% across a 42-variant grid, with about 52% of in-sample Sharpe surviving out-of-sample. The figure is window-dependent, and the report documents that instability rather than quoting the friendlier run. 195 tests in CI.
Tradeski · tradeski.dev — Live market, macro, and sentiment data in one dashboard: 11 ticker symbols, 7 FRED series hourly, a 29-symbol screener, a retrieval-grounded assistant, 121 tests. Its most useful lesson was a bug: live prices never reached the browser because the server emitted on one Socket.IO namespace while the client listened on another. Both halves were correct alone, and all 71 tests passed throughout — every one of them tested a pure function, so no test covered a contract between components. One now does.
Neural Canvas · github.com/Danny-397/neural-canvas — Painting by hand gesture through a webcam: 21 3-D landmarks per hand at ~60 inferences/second, fully on-device, one HTML file, no build step, no server. The hard problem isn't detection but estimating intent from a noisy 60 Hz signal — a fixed low-pass filter trades jitter against lag, so strokes use a one-euro adaptive filter whose cutoff rises with hand velocity.