Trippie: a favorite toy, a new purpose.
Meta Global AI Developer Hackathon · Indie / Push the Frontier entry
What
A neighborhood inspection prototype that turns street-level photos into located, highlighted evidence and draft city maintenance cases. A Pixel phone mounted on a favorite toy robot gives Trippie its point of view.
- Verified field trial: 15 Pixel photos, two draft cases, and three reviewed SAM highlights.
- Original photos, highlighted regions, AI interpretation, and case review in one portal.
Why
“It was a terrible fall. I know—because it was me.” A fall in my neighborhood inspired Trippie: notice the hazard before the next person trips.
A study of 2019 U.S. EMS records identified 118,520 injurious falls on streets and sidewalks—not all caused by damaged pavement. Study ↗
How it works, step by step
Follow the verified photo-to-draft workflow and the technology used at each step.
- 1
Capture the evidence
Pixel 10 camera · JPEG / EXIF · local intake
Street-level photos supply the original evidence, capture time, and available GPS. The field trial uses 15 imported Pixel photos. Location accuracy is recorded when available; an unreported value stays unknown.
Produces: Original photo + capture metadata
- 2
Interpret the scene
Muse Spark Open · Fireworks
Muse identifies visible maintenance concerns, records observed facts and uncertainty, and suggests a city form and draft description. Photos with no actionable finding stay in the observation feed.
Produces: Structured observation + model request ID
- 3
Locate and group the finding
City of Plano GIS · Muse multi-image reasoning
Photo GPS is matched to nearby road geometry and connected intersections. Muse compares multiple views to group one physical issue into one case. Locations remain estimates, with intersection offsets; reused GPS alone does not establish a duplicate.
Produces: Estimated intersection + grouped evidence
- 4
Highlight the problem area
SAM 3.1 · Meta Model API · mask review
SAM uses region, box, and point prompts to segment the selected surface or object. Each published highlight is checked against the source photo, request, prompt and image hash. The portal preserves the original alongside the reviewed mask.
Produces: Original + reviewed SAM highlight
- 5
Prepare the city case
Audited Fix It Plano forms · draft case payload
The draft contains the request type, estimated location, concise description and original-photo attachment. The field trial produces two draft cases. Form preparation was checked in the browser; no city request was submitted.
Produces: Draft case ready for final form review
- 6
Publish and inspect
Supabase · Next.js · React · Leaflet / OpenStreetMap
Supabase stores observations, case records and image assets. The public portal shows the map, original/SAM comparison, magnifier and case feed without sign-in. Protected operator commands manage capture mode and disposable DEMO data.
Produces: Public evidence portal + durable case records
Architecture · Mermaid
Photo capture, model interpretation, checked evidence, public review.
Rendering architecture diagram…
Mermaid source
flowchart TB
accTitle: Trippie verified photo-to-draft workflow
accDescr: Imported Pixel photos are interpreted by Muse Spark Open on Fireworks. Plano GIS estimates location and Muse groups views of one issue. Reviewed SAM 3.1 highlights and draft city cases are stored in Supabase and displayed in the public Next.js portal. Observations with no actionable issue are retained without creating a case. City submission is outside this workflow.
A["1 · CAPTURE
Pixel photo + EXIF"] --> B["2 · INTERPRET
Muse Spark Open on Fireworks"]
B --> C["3 · LOCATE & GROUP
Plano GIS + Muse multi-image reasoning"]
C --> D["4 · HIGHLIGHT & CHECK
SAM 3.1 via Meta Model API"]
D --> E["5 · PREPARE
Draft city case + original attachment"]
B -->|"No actionable finding"| O["Observation only"]
E --> S["SUPABASE
Postgres records + image storage"]
O --> S
S --> P["6 · PUBLISH & INSPECT
Next.js portal + OpenStreetMap"]
classDef model fill:#e5efd5,stroke:#9ab67c,color:#344d2b
classDef local fill:#f7f9f1,stroke:#c2d0b0,color:#415837
class B,C,D model
class A,E,O,S,P local
The diagram follows the implemented imported-photo workflow. SAM marks visible regions; the photos do not establish defect height. Draft cases are distinct from submitted city requests.
What matters technically
- Open-weight reasoning: Muse Spark Open is hosted on Fireworks; it runs in the cloud, not on the phone.
- Hardware baseline: the replacement Rugged Robot travelled five forward steps straight through its official app. That bench test is separate from the photo-analysis workflow shown here.
- Evidence integrity: keep original + SAM output, timestamps, location accuracy, and model request IDs with every case.
- Original work: new portal and inspection workflow. Existing toy hardware and earlier Pi experiments are the starting point.
Judging alignment
- Impact · 30% Earlier hazard discovery and actionable evidence for city maintenance teams.
- Meta tooling · 25% Muse Spark Open for reasoning; SAM 3.1 for segmentation.
- Technical execution · 20% Traceable model outputs, verified image migration, durable storage and a working public evidence portal.
- Demo · 25% Follow one finding from camera capture through model reasoning, SAM highlighting, and a prepared city case.
Weights: participant guide reviewed Oct. 31. Guide ↗