Weed Scout

Weed Scout: upload the photos from one walk through a corn, soybean or alfalfa field and get a spray decision grounded in the herbicide label. Whether a field needs a herbicide pass is still decided by walking it and judging by eye. A scout has to tell weeds from crop, name the species (one resistant waterhemp or Palmer amaranth changes the call), judge how many there are and how big, compare that with thresholds, then find a product whose EPA label covers this crop, this weed and this weed height. I've worked in agriculture and know row-crop scouting firsthand: it is slow, and mistakes cost yield or waste sprays that build resistance. For every photo in the walk: 1. Survey: a Muse Spark agent says which crop the photo shows and about how many weeds it sees, and flags a crop that doesn't match the walk. 2. Mark: the agent boxes the crop rows and every weed it can tell apart (up to 120 boxes in one call); SAM 3.1 turns the boxes into masks, 16 per request, in parallel. Where weeds are too dense or small to box, it draws a weed area that code measures from green pixels minus crop. 3. Check: after each mark the agent sees a picture of what it has marked, compares it with the photo, fixes what it missed and zooms into doubtful plants. 4. Identify: for each kind of weed, its clearest plants go to a panel of seven independent Muse Spark calls (a full-resolution leaf close-up plus a wider view). Each describes leaf shape, margin, veins, hairs, stem and habit before naming the species, and the panel votes. Naming is open: a weed off the catalog is named in its own right instead of being forced onto the nearest listed weed. 5. Measure: code, never the model, computes weed cover, counts, density and in-row vs between-row share. 6. Decide: rules compare species with sourced thresholds (zero tolerance for resistant amaranths) and choose no action, scout again, or treat (spot or broadcast). 7. Ground in labels: a table of 1,680 use records extracted from 20 EPA label PDFs, each quote verified on its cited page, lists the products labeled for this crop, weed and weed size. You watch the agent work live: each marking picture, close-up, final overlay and its own reasoning. The report has a decision card, per-stop overlays, a species table and label-backed options with page citations. Two sample walks replay saved runs instantly. Live: https://weed-scout-939814041857.us-central1.run.app Built entirely during the hackathon. Decision support, not a prescription: the label is the law.