Concourse

Concourse escorts a parent who cannot use apps or read English through an international flight — on WhatsApp, by voice, in Telugu, with her family in the loop. My mother can fly 9,000 miles alone. The part she cannot do is read the airport. Every stage of a HYD → DOH → DFW trip — check-in, security, gate, boarding, the Doha transfer, immigration, baggage, exit — arrives on her phone at the right time as a Telugu voice note and a card she can match against the boards (the board word in English capitals, huge, plus the gate code). She answers the only way she knows: a voice note or a photo of a sign. Concourse answers only from the verified itinerary and what is actually printed in her photo. A gate change reaches both phones. One tap shows a card to airport staff; one tap tells her son "I'm here." Built on Meta — every model that runs is a Meta model: • Muse Spark 1.3 reads the ticket into a schema-constrained itinerary, classifies what she means (intent · confidence · tool proposal), and reads the signs she photographs (location_claim is typed null — the model cannot say where she is). • SAM 3.1 finds the sign in every photo and camera frame, in parallel with Spark — a highlight, never a route. • Muse Voice Transcribe 1.0 transcribes her Telugu voice notes (Telugu/English bias, airport vocabulary). • SeamlessM4T v2 (Meta open weights, our Modal deployment) is her voice; MMS-TTS Telugu is the fallback. • WhatsApp Business Cloud API is the channel: voice notes, cards, three-button replies, a link to her travel screen (card, voice, live camera reading of the boards, hold-to-talk). What is deliberately not a model: the journey state machine and its timing, the gate (only a verified operations event changes it), the safety check on every outbound message, and the cards. Spark interprets the human; code controls the journey; Concourse never invents a direction. The safety contract is enforced in code and tested by 36 adversarial cases in CI: no location claims, no invented routes or gates, a photo never overwrites the itinerary, printed text is never obeyed, immigration questions are never answered for her, she is never told to ignore staff. Evidence: 36/36 safety cases green; the full demo path (15 stages, gate change, 4 sign photos, 3 voice notes, the ticket) run end to end against live Meta model calls and replayed by CI on every commit: 103/103 model calls, 7/7 of her inputs answered from verified facts, 0 invented directions. Not verified before the deadline, said plainly: the journey on real WhatsApp phones (the client is built and tested against the API contract; our test business account was locked), and a walkthrough with a real elderly user. The video is the real app replaying its recorded model calls with her synthesized voice — nothing is staged outside the code. Run it: the README's five-minute path replays the recorded run on a laptop with no keys; with a Meta API key the calls go live. Built Oct 26 – Nov 4, 2026. Apache-2.0.