# RiverEye - **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon) - **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST) - **Where:** Online - **Team:** [Matt Brady](https://cerebralvalley.ai/u/mattbrady) - **GitHub:** https://github.com/by42ppcrps-dev/rivereye - **Demo video:** https://youtu.be/XlaV6290VYs - **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery - **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1106 RiverEye turns drone video of a canal into a streamflow measurement you can check. Irrigation districts and small utilities measure flow a few times a year with a crew and a current meter. A drone can fly a canal in minutes, but video alone gives a number nobody can verify. Meta SAM 3.1 finds the water and tracks hundreds of foam flecks. Discharge is a depth factor times their speed times the cross section USGS surveyed. Muse Spark picks the depth factor without seeing the reference flow, audits 24 tracks and writes a field note in English and Spanish. Muse Voice Transcribe turns a spoken question into text so a manager can ask the measurement about itself. Today that runs from the command line on a recorded WAV, and the web page replays two recorded questions. Any note or answer with a number not in the measurement is sent back. On the public USGS Coachella Canal clip it reads 19.7 m3/s against 19.5 from the current meter, 0.6 percent high. A method with no AI in the speed measurement agrees on speed, 0.89 against 0.91 m/s. At the common depth factor of 0.85 the error would be minus 4.9 percent, so the app shows a range and an alpha slider. It fails on our second site, the Cochiti canal, where foam is sparse. It reads 31.9 percent low, and Spark's depth factor made it worse. Without using the USGS number, RiverEye marks that run low confidence because only 13 percent of tracks survived the filters. We set that rule after seeing both results. We then ran the unchanged pipeline on 8 more USGS sites with a 15 percent pass rule set before the runs: 4 of 10 pass and 6 fail. Cochiti reads low, and 5 sites where nothing drifts steadily with the flow give no estimate. The flag said low on almost every run, so it is a warning, not yet a filter. We kept testing after the demo, with the shipped default unchanged. With an opt-in corner tracker (inside SAM's water mask) and an opt-in depth factor computed from the Froude number, with no model call, RiverEye passed 8 of 10 USGS sites in leave-one-out scoring (6 for a fixed 0.85) and 3 of 4 sites from a second public dataset (2 for 0.85). Muse Spark's pick on the same tracker passed 7 of 14 in total. The new-site gain is one borderline site (14.97 percent low against a 15 percent line), and we chose the options with results in view, so we call it promising, not proven. With those options RiverEye also gives an 80 percent range of about 15 percent below to 18 percent above (it held on 10 of 12 sites) and declines to estimate when nothing drifts steadily. Results, caveats and one deviation from our own protocol are in the repo. ## More from Meta Global AI Developer Hackathon - [Jeevana sati](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1103) - [LEPCPRO](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1104) - [Meqyas AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1105) - [Crisis](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1107) - [ADADIVE](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1108) - [PrinceDevs](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1109) --- Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1106. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt