Ayush Kumar (solo)

At our college gate in Bhopal, a student and a motorcycle reached the same spot at the same instant, 0.42 m apart. Crossing Watch measured it from one phone video. The problem: road crashes kill about 1.19 million people a year and are the leading cause of death for ages 5 to 29 (WHO, 2023). In India they killed 35,221 pedestrians in 2023, and two-wheelers killed more of them than any other vehicle (MoRTH, Road Accidents in India 2023); all 5 near-misses at our gate were with a motorcycle. Every school wants a zebra crossing at its gate. Only the road authority can build one, and it asks for proof. What it does: Meta SAM 3.1 tracks people and vehicles. Four marked corners on the road turn pixels into metres, and code measures each near-miss (post-encroachment time and vehicle speed). Muse Spark looks at every tracked "person" in a near-miss and sets aside riders and empty boxes, then writes the letter. Code checks every number in it and two AI reviewers read it. Muse Image draws the requested fix only on the strip where people crossed. A case agent stops and asks a person to tape-measure a line it never calibrated from. Faces and number plates are hidden on every published frame, and the site works in Hindi and English. Our gate, 8.6 minutes of dismissal: 55 moving vehicles, 80 tracked crossings, 51 people on foot. Muse Spark set aside 66 of 75 alarms and a human check of the remaining 9 kept 5 real near-misses: 2 serious (0.0 s and 0.9 s) and 3 close passes. The video measured 15.0 m where the tape read 15 m. The letter asks for a painted zebra and "College Zone 20 km/h" signs, and on 3 Nov 2026 the college certified the footage and report as genuine. Its Head of CSE (Data Science), Dr. Neetesh Kumar Gupta, writes (4 Nov 2026): "We have had several close calls, and it has been a persistent concern for the department and the college administration. Until now, we did not have any documented evidence to formally highlight this issue to the concerned road authorities." Proof beyond our gate: 47 of 52 near-misses found against drone ground truth (DUT), 243 offline self-checks, and the whole gate case also ran with every reasoning step on Meta's open weights. Open weights: on SummEdits, a public benchmark labelled by people, our LoRA-tuned Muse Glimmer judge beats Muse Spark Open, 1,008 vs 924 of 1,200 (p < 0.0001, pre-registered); base Glimmer gets 748. Try it in a browser, no key: https://crossingwatch.xyz/app/ For judges, a 5-minute map: https://github.com/shaunmortal/crossing-watch/blob/HEAD/JUDGES.md Open weights, one page: https://github.com/shaunmortal/crossing-watch/blob/HEAD/OPEN_WEIGHTS.md Tuned judge adapter: https://huggingface.co/shaunmortal/crossing-watch-glimmer-judge-v6 Solo entry by Ayush Kumar, Data Science student at LNCT Excellence, Bhopal, India; built with Claude Code as a coding assistant (declared in the README).