Tessa

Tessa is an AI agent that hunts for planets around nearby red stars in public NASA TESS data, and accelerates a real scientific search. TESS produces far more repeating dips than people can check, and most are fakes: eclipsing binaries, neighbouring stars, spots or noise. From a pool of 99,461 K and M dwarfs within 100 parsecs, Tessa searched 76,318 stars, every one that never had 2-minute TESS data and has a usable light curve, plus 27,552 two-minute targets with at least two more full-frame sectors, and found 2,698 repeating dips (about 750 is what noise alone gives, measured by searching the same light curves upside down). A fine-tuned Meta model investigated the 715 that code did not settle on its own, and 8 new candidates passed Tessa's checks and the bar of TRICERATOPS, a false-positive calculator: 4 in its "likely planet" class, and 4 where a nearby star is not yet ruled out. Three are on no published list we checked; other searches found the other five too, and Tessa reached each independently. Tessa also re-found, blind, the one candidate our hand-written scripts found before the hackathon (TIC 458191673, published as a preprint), which is not counted among the 8. On 1,000 stars held back before any search (999 with usable light curves) and searched once with the pipeline frozen, it made no candidates out of noise, and it still makes none under the final gate 2. Every dip gets a written case file, and every candidate gets a dated prediction of its next transit, from the ground or from TESS, that anyone can test. A candidate is not a planet until a telescope confirms it. Everything is on a live 3D map at https://tessa.braveai.solutions, where Ask Tessa answers questions from the survey's own files.