Suri Project

Crash records only see the collisions that already happened. At the García Roel and Ahuehuetes crossing in Monterrey, Mexico, INEGI open data records 16 crashes in 2019–2024 and no pedestrian hit. Near misses never reach a database, yet they are where the next injury comes from. Suri turns one fixed camera into evidence of those close calls: - SAM 3.1 tracks pedestrians and cars from a text prompt, with no training and no manual labels. - Code measures post-encroachment time (PET): the seconds between one road user leaving a conflict zone and the other entering it. Under 3 s is a close call; under 1 s it is serious. - A Muse Wonder agent verifies each candidate with tools: it calls a kinematics tool, asks Muse Spark to watch the blurred clip, and submits a verdict that code re-checks against fixed rules. - The output is an inspection report in Spanish for traffic engineers. Code fills in every number; people and license plates are blurred. Our rule: the code computes, the AI explains, the engineer decides. Demo run: on an open roadside-camera clip from the Universidad de Murcia (CC BY 4.0, 2.7 min), code found 7 candidate close calls. In two tool-calling turns, Muse Wonder verified 4 (3 serious, 1 close) and rejected 3 whose paths never met inside the event window; code re-checked every verdict. Recording the Monterrey crossing is the next step.