You don't learn what risk feels like by reading about it

Every financial literacy app I tried was a textbook with a quiz bolted on. Capital Crusade runs ten simulated market years in a few minutes on a phone: you allocate across equities, debt, gold, cash and real estate while the prices move underneath you — stochastic drift, per-asset volatility and a correlation matrix, so the assets move together the way real ones do. Macro events fire off state triggers and probability distributions rather than a fixed script, and the end-of-run debrief scores risk-adjusted performance rather than raw return. Getting lucky and being right shouldn't look the same.

A phishing detector that tells you why

Spam filters fail in two directions and the second is worse — flag someone's real bank alert twice and they stop trusting the filter forever. PhishGuard runs four layers: a TF-IDF and logistic regression baseline, 24 structural features the model can't see, a structural override for adversarial emails too short to classify, and safe-indicator reduction that looks for the trust signals phishing never bothers to fake. Every verdict lists the signals that fired. Ships as a Chrome extension that injects into Gmail.

Known limitation: very short phishing with no suspicious TLD can slip through. The model needs two structural indicators to override a low ML score.

Flood monitoring for field agents with no signal

Rivers get monitored where connectivity is worst, which is exactly where a normal web app falls over. Offline-first PWA: capture the reading and a photo of the gauge with no internet, persist to IndexedDB, sync on reconnect. Gemini Vision reads the staff gauge from the photo and validates the scene — it catches someone photographing a phone screen instead of a real gauge. RandomForest risk classification paired with a Manning's equation engine that generates flood extent polygons.

Dictation that never leaves the laptop

The good dictation tools send your voice to somebody else's server, which is a strange trade for something you talk to all day. Parakeet-TDT does the transcription on the Apple Neural Engine through MLX, and a local qwen2.5:7b through Ollama runs a second pass over the raw transcript: strips filler, fixes punctuation, and resolves spoken self-corrections, so "meet at 2, actually make it 3" comes out as 3:00 PM. That second pass is the difference between dictation you keep and dictation you rewrite. A background daemon holds a global hotkey — hold right Option anywhere in macOS, talk, release, and the cleaned text pastes at the cursor.