See what the model sees
Sovereign Shield 0.6.0 is live: an Inspect panel showing your prompt beside what the provider actually gets, plus API-key detection, custom terms and realistic stand-ins. Still entirely local.
Archive
Projects I’ve retired, and the posts I wrote while they were alive. Everything here is kept as it was written; it just no longer sits in the main list.
10 posts · current writing
A deterministic privacy shield for LLM prompts — a Chrome extension and a Python proxy. Frozen in September 2026: the extension stays on the Web Store as it is, with no further updates.
Why it stopped →Sovereign Shield 0.6.0 is live: an Inspect panel showing your prompt beside what the provider actually gets, plus API-key detection, custom terms and realistic stand-ins. Still entirely local.
Sovereign Shield's browser extension is live on the Chrome Web Store: it redacts Swiss, EU and international identifiers before your prompt reaches ChatGPT, Gemini or Claude.
A deterministic gateway that lets your team use Gemini or Claude without sending Swiss personal data across the border — tokenised on the way out, restored on the way back.
An autonomous AI red-teaming framework — attacker, target and judge models pressure-testing where frontier models leak. Research concluded in 2026; the findings stay up as written.
Why it stopped →I re-ran the same secret-extraction attacks against the flagship models. Gemini 2.5 Pro folds to a forced schema 95% of the time — worse than the cheap tier — while Claude's flagship models drop the prefill surface entirely. The holes move between generations.
Gemini leaks under a forced JSON schema; Claude resists that but folds to prefill. Opposite weak spots, same secret — and one deterministic defense that catches both.
Six hand-written jailbreaks couldn't crack my hardened bot. Then I forced its reply into a JSON schema and it leaked the secret on turn one. Why structured output is the hole.
I built K.E.V.I.N.: one AI tries to extract a secret from another while three judges score each attempt. Three things 50 rounds of automated jailbreaking taught me.
A WebGL visualizer that turned a track into a 3D “sonic fingerprint.” Retired in June 2026; its audio analysis lives on inside 1-bit.
Why it stopped →Two weeks, three new music visualizers, and one absurdly satisfying landscape. Here's what changed in Sonogram — generative audio art for your eyeballs.
A photorealistic turntable, a snowy Alps landscape, a brand-new About page, ~130 commits — and a deep rewrite for frame-rate independence in Sonogram.
A deep dive into Sonogram, a full-stack web app that turns a song's mathematical properties into an interactive, shareable 3D 'sonic fingerprint'.