I'm a developer and PhD student at the Universitat de València, in the NIMLab group, where my research is on immersive virtual reality experiences for cultural heritage. I make apps for the web and the Mac, and tools for coding agents. My site is nerln.github.io.
|
|
|
|
|
|
I build tools, then I publish the thing that checks them.
When a claim can be checked, the check ships with the code. plancia runs 4956 checks before every push, against a throwaway archive that never touches yours, and its release script refuses to build a release if any of them fails. claude-codex-bridge writes a session only for CLI versions it has been verified against, and ships the drift check to run after either CLI updates.
That is the whole idea. Anybody can claim a tool works. The interesting part is handing you the thing that would catch me if it did not.
The clearest example is not a tool. placa forecasts a reality show from the vote splits the broadcaster reads out on air. It also refits itself backwards, each gala using only the galas before it, and reports the result on the page even though it is bad: over the eight galas it could test, up to 14 September 2026, its ranking put the person who actually left 4.88th on average when chance would put them 3.44th. Zero hits out of eight, p = 0.0091 in a one-sided permutation test. The test is on the page, and the data is in data/retro.json of the repository.