[01] Vessel performance — pipeline & physics-hybrid ML

Core stages of the central data pipeline: multi-vendor sensor ingestion, formula-driven measured columns, timestamp-aware position resampling.

A PyTorch architecture wrapping a hydrodynamic resistance simulation and learning the residual, plus the evaluation standard around it — voyage-level holdouts, autocorrelation-aware splits, leakage detection, experiment tracking. Also a multi-provider hindcast weather pipeline and a headless sea-route generator feeding the company API.

[02] Digital-twin voyage simulator

Photorealistic Unreal Engine 5 simulator, modelled in Blender, replaying real routes and real weather with measured and predicted power side by side along the track. Built unprompted; now used in customer demos.

[03] Quantum-inspired probing of neural representations

Reflection-based probing of hidden activations, examining what a network's internal states actually encode.

[04] News sniper on prediction markets

Does a fast news feed plus a fast typed-judgment model beat a prediction market's reaction? Wire posts from 28 news accounts arrive over the Bluesky firehose about two seconds after publication. Each one goes to Jev (TypeSafe's System One model) in two calls: which of ~600 Polymarket markets does this move, then a read of that market's resolution rules for direction, fair probability and confidence. Trades are simulated at the real order-book fill and scored one hour later and at resolution.

Honest status: the experiment exists to find out whether the market has already moved by the time the feed shows the news. First days say: RSS is minutes late and useless; the firehose is fast enough to make the question real.

[05] Outreach pipeline for one-off data jobs

Finds small firms that publish messy structured data (listings, vacancies, tariffs), crawls the page, extracts every record verbatim, and lets plain code find the mess: duplicates, three different price units, listings still shown as rented. DeepSeek writes a mail under 120 words in the recipient's language; Jev gates it on six checks and a hard rule that every number must exist in the extracted rows. A human approves each send. First run: 22 companies in, a handful of sendable drafts out, the rest correctly rejected for having nothing to fix.

[06] Jeb — voice control that acts before the sentence ends

Offline speech recognition streams partial transcripts every 50 ms; each partial goes to Jev in one fan-out request (command, target app, amount, "has enough been said?"). Code acts the moment confidence clears a per-command threshold: "open spot…" and Spotify is opening, 400 to 800 ms after the first word, usually before the sentence ends. Wake word, 31 apps, media keys, screenshots, lock.

[07] Paper-trading fleets

Hundreds of rule-based and ML bots on crypto, gold, Nasdaq and FX, paper-traded on real market data with the same decision function the backtester uses. Walk-forward validation, a leverage ladder that replays every bot at 1x to 50x until liquidation, and an honest scoreboard: in-sample 58 to 79 percent, out-of-sample 49 to 54 against a 46 to 51 base rate. The result is the point.

[08] GMSR Studio

Small objects that carry a feeling from one person to another. First object: Limonchito, a squeeze-to-buzz pair. Blender model, KiCad schematic, firmware, production plan, and a 3D room to walk through.