PATROL. Slides · 16
  1. Slide 1: MULTI MODEL HACKATHON · SF TECH WEEK PATROL A search drone that looks for one missing person’s things, with no internet. “The camper had a blue backpack, an orange tent, and a gray dog.” 3 of 3 things found on a real-time flight, no false alerts 0.32 s median time to judge one odd-looking spot Simulated flight, drawn from its own log 1
    01 / 16
  2. Slide 2: Everything it sees, thinks and finds 1 2 3 4 5 6 7 8 1 Header state, area searched, battery, who decided 2 Looking for the clues, in qwen38’s words 3 Camera the sweep photo, then the close-up and verdict 4 Decision trace the crop Clef-Flash judged, and its odds 5 Map lanes, trail, spots checked, alerts, next sectors 6 Chase view for the audience; no model uses it 7 Log every decision, not
    02 / 16
  3. Slide 3: Watch the whole flight 3 min 23 s, replayed at 3.5× with the alerts read aloud. Three finds, the tarp checked and dismissed, no false alerts. This recording is also the stage fallback. 3
    03 / 16
  4. Slide 4: Searchers start with words, not photos WHAT THE FAMILY SAYS “The camper had a blue backpack, an orange tent, and a gray dog.” Nobody has trained a detector on this person’s blue backpack. No cell signal: an alert is about 270 bytes of text plus a 20 KB photo. Decide in the air. After landing, every “maybe” costs another flight. From 40 m, a backpack is a smudge About 22 pixels across: eno
    04 / 16
  5. Slide 5: A flight in five moments “...a blue backpack, an orange tent, and a gray dog.” 1 Say it Parakeet writes it down. qwen38 turns each thing into a sentence about how it looks from above. 2 Sweep at 40 m A photo every 12 m. The finder boxes odd spots; Clef-Flash says 61% backpack in about 0.3 s. 3 Follow a lead Down to 20 m for a close-up. The backpack scores 94.5%: confirmed. ALERT 4 Alert A
    05 / 16
  6. Slide 6: A decoy dismissed, a backpack confirmed FROM 40 M CLOSE-UP AT 20 M DISMISSED FROM 40 M 17% blue backpack? AT 20 M 4% A flat sheet. Best clue score 4%. FROM 40 M CLOSE-UP AT 20 M CONFIRMED FROM 40 M 61% blue backpack? AT 20 M 94.5% Straps and pockets. Back in the lane with an alert. Scores are Clef-Flash’s probability that the spot is the blue backpack. 6
    06 / 16
  7. Slide 7: Four models fly it, three built the props Clef-Flash THE FAST EYES Judges each odd spot in about 0.3 s qwen38 THE SLOW THINKER Clue sentences, radio notes, trail plans, second opinions Parakeet THE EARS Speech to text in 20 to 90 ms Kokoro THE VOICE Reads every alert aloud Plain code flies the drone and owns safety. Models advise; no model can override it. THREE MORE MODELS, CHAINED BY m3
    07 / 16
  8. Slide 8: Fast loop, slow thinker, code in charge Clef-Flash judges every photo in about 0.3 s. qwen38 does a few background jobs. Code checks every answer and owns safety. 8
    08 / 16
  9. Slide 9: Fly high, look close sweep · 60 m the finder misses it sweep · 40 m the sweep height sweep · 20 m easy, but twice the lanes zoom · 20 m the first closer look zoom · 12 m only if unsure 1 of 10 views where the finder found the backpack from 60 m 10 of 10 views where it found it from 40 m, so 40 m is the sweep height 7 vs 13 lanes to cover the 160 m square from 40 m versus 20 m 9
    09 / 16
  10. Slide 10: Higher means fewer lanes 40 m 20 m 12 m 0 m two neighbouring lanes, seen end-on 22.9 m between lanes overlap 32.6 m of ground in each photo a closer look at one spot first look at 20 m second step at 12 m, only if unsure hover 0.6 × the height short of the spot, about 31° off vertical; the zoom frames about 4 m of ground backpack, 0.55 m (drawn to scale) At 40 m, 7 lanes cover the square
    10 / 16
  11. Slide 11: Each model sees only what it must One flight: 683 s over a 160 × 160 m square. Clef-Flash’s 171 calls took about 56 s of model time; qwen38’s 6 jobs about 15 s, in the background. 375 odd spots boxed by the CPU finder, on every photo 163 crops judged by Clef-Flash, plus 8 close-ups 6 qwen38 jobs: 3 radio notes, 3 trail plans 11
    11 / 16
  12. Slide 12: Talking to it THE OPERATOR SAYS CLEF-FLASH ROUTES IT TO WHAT HAPPENS “Missing hiker. Blue backpack, orange tent, grey dog.” describe qwen38 writes the clue cards “Start the search.” start_search take off and sweep “Go back to the last alert.” go_back_to_last_alert fly back and look again “Stop searching and come home.” return_home straight home 20 to 90 ms Parakeet, speech to text 45% mi
    12 / 16
  13. Slide 13: One laptop, no internet 128 GB One MacBook Pro M5 Max A 40-core GPU and 128 GB of unified memory run every model, the simulator and the console at once. No internet Everything is local Every model runs on the laptop and nothing in the code calls out. 0.3 to 0.4 s One GPU is the limit Clef-Flash answers that fast when the GPU is free. With another job running, 2 to 6 s. 13
    13 / 16
  14. Slide 14: It works end to end, in a simulator 3 of 3 found on a real-time flight, 683 s, no false alerts 17 of 18 scenario flights pass, over three sets of fresh seeds 0 in 45 false alerts across 45 flights and 95 alerts 0 vs 7 false alerts, live versus reviewing photos after landing Found from words alone Dog, helmet (0.3 m) and sleeping bag: items never used for tuning. Read these as “this behav
    14 / 16
  15. Slide 15: What doesn’t work yet PASSES, ONE MISS One miss left: the finder On a 30-minute battery the four-clue flight covers the whole area; 17 of 18 pass. The miss: the finder never boxed one sleeping bag. PARTLY New items lean on qwen38 Clef-Flash’s close-ups score 0.12 to 0.29 on held-out items. Two of four finds came via a second opinion. OPEN It has only seen a simulator Flat ground, no wind
    15 / 16
  16. Slide 16: What’s next 1 Fix the finder's blind spot The last held-out miss was a sleeping bag the colour finder never boxed. 2 Wi-Fi-off rehearsal Run the whole demo with the network actually off. About ten minutes. 3 Real aerial photos Replay HERIDAL, SARD or WiSARD imagery. Approved; dataset licences are being checked. 4 Hybrid after-landing pass Re-check saved photos, but only send a team to sp
    16 / 16