# Applanta Solutions - **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:** [Rudy Miller](https://cerebralvalley.ai/u/applanta_dev) - **GitHub:** https://github.com/ApplantaSolutions/ridertap-ai-dismissal-copilot - **Demo video:** https://www.youtube.com/watch?v=FM1AubtftuA - **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery - **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/607 School dismissal is the one part of the school day still run on memory and walkie-talkies. Staff at the curb shout car-tag numbers down a line; staff inside send children out; nobody is actually watching the loading zone to check the cars physically in it against the wave that was called. When a car is skipped or arrives out of order, the problem is discovered at the curb, with a child already outside and the line stopped behind it. The context is not abstract: - 54.2% of children ages 5–17 are typically driven to school in a private vehicle — more than take the bus (2017 National Household Travel Survey, via U.S. DOE). - Schools employ 21,200 fewer bus drivers than in 2019 — a 9.5% decline (Economic Policy Institute). - 81% of school administrators say pickup and drop-off congestion is a problem, and 83% say staff get pulled from their core jobs to cover transportation (AP-NORC / HopSkipDrive 2025, n = 510). - The largest share of weekday child pedestrian deaths — 28% — falls between 3 and 6 p.m. (NHTSA, Children 2023). - In Washington state, a Traffic Safety Commission radar study of 118 schools (December 2023) found nearly 3 of 4 drivers exceeding 20 mph in school zones. And at the founder's daughter's school, the official car-rider window is 2:00–2:15 p.m. — yet parents queue from about 1:30, and by 1:49 on the day we documented, both lanes were full and the queue had spilled onto the public road, eleven minutes before dismissal began. The founder sits in that line most afternoons; that account is firsthand. **RiderTap** is the base product: a live dismissal board (in production since October 24, 2026) where staff tap a rider number and every screen on the line updates instantly. No names, no rosters, no photos — numbers only. The **RiderTap AI Dismissal Copilot**, built during this event, rests on one insight: every dismissal tool manages what staff *report* — the Copilot checks the physical loading zone against the call. It adds three capabilities, each carried by a different Meta model: 1. **Zone check (SAM 3.1).** A staff phone framed on the loading zone takes a still every few seconds and crops it to the zone on-device; only the crop is sent for counting. If the count disagrees with the called wave on two consecutive stills, the board flags it — a skipped or extra car is caught at the zone, not at the curb. The frame is discarded the moment it is counted; only the count is kept. 2. **Exception agent (Muse Spark, standard model ID).** When anything goes off-script — a zone mismatch, a duplicate number, a safety hold, an out-of-order arrival, a walk-up — the agent reads the live board state and gives the lead one plain-language instruction (140 characters or fewer, with a structured action). It is advisory only: a human confirms every action. It runs against a 3-second hard cap with a rule-based fallback behind it, so the line never waits on the model. 3. **Voice entry (Muse Voice Transcribe).** The staff member walking the queue holds a push-to-talk button and speaks tag numbers instead of typing them. The transcript may only resolve to a tag actually in the live queue. If two in-queue tags sound alike, the number is read back for confirmation before anything is entered. On unclear audio it asks for a repeat; it never guesses, and it never enters a number twice. Audio is discarded after transcription. Each of the three is load-bearing — remove any one and a named capability dies with it. A replay mode runs recorded zone stills, voice clips, and scripted board events through the same live API calls — the exact demo storyline can be re-run with model calls still live; the final build passes 25 of 25 runs. **For judges — run it live (60 seconds).** The public repo's README opens with a step-by-step tour: the live demo link, join code RT-KW4W4X, and the film's wave already loaded on the board. On the Walker tab, say "three four five" and the voice entry stops at a read-back confirmation instead of guessing; on the Zone phone tab, "Use sample lot photo" runs the zone check against a lot photo and raises the mismatch the exception agent then answers. **Measured results** (full table in the public repo). Across two 1,000-call batteries through the deployed app — 2,000+ live model calls — zero wrong entries, zero silent entries, zero invalid instructions. - **SAM 3.1** — zone counts exact in all 493 answered checks across both batteries, with the only non-answers failing closed rather than miscounting. - **Muse Spark** — every staged exception event in both batteries produced a valid, safe instruction inside the 3-second cap; the model itself served about 88% of them, the rest by the designed rule fallback. - **Muse Voice Transcribe** — zero wrong entries across 250 hostile voice cases and 62 twin-flip attempts on the hardest sound-alike clip; every mishear was caught at read-back confirmation or failed safe by asking again. Pilot discussions are underway with Barrow County schools. One principal who reviewed the overview and walkthrough said, "I think you're on to something." No school is named here, because naming permission has not been granted — which is the same discipline the product itself is built on. **Honest limitations.** Muse Spark's latency on zone-mismatch prompts sits at the design edge — median about 2.5 seconds against our 3-second cap — so the rule-based fallback is a permanent part of the design, not an apology: we do not claim the model serves every event. The transcription model can mishear sound-alike tags: in an earlier hostile battery (hurried speech, lot noise), exactly one wrong entry in 92 utterances occurred — a tag one digit off from another live in the same wave. The guard now fires read-back confirmation for any confusable tag in the wave — the exact clip behind that entry stops at confirmation on the deployed build, and both 1,000-call batteries since show zero wrong entries. Production RiderTap is untouched by all of the above. All dismissal testing so far is staged and simulated — real lot noise, weather, and school Wi-Fi are what a live pilot is for, and we have not run one yet. ## More from Meta Global AI Developer Hackathon - [Chumba Labs](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/604) - [Harbor Muse](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/605) - [DeshkaAI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/606) - [Tengacion Technologies Limited](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/608) - [FractalSwarm](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/609) - [Eliobros Tech](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/610) --- Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/607. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt