Idea

A model describes. Code decides.

SDOT camera frame
Earlier read of this frame: Qwen2.5-VL, safe-walk schema
3/23
Cosmos-Reason1-7B gave a safety verdict
0/23
qwen3-vl:8b, same prompt, same frames

Same 23 frames as these thumbnails MEASURED GB10, Aug 2026, vlm/RESULTS.md §3. Reads: lab/samples/mac_reads.json.

Me: vlm, audio-lm, ios-voice, calling Teammates: media-ingest, pathfinding, walk-app, osint Dhruv: model endpoints on the Spark Claude Code (AI): co-author on all 78 of my commits

NVIDIA Spark Hack Seattle, Aug 14–16 2026, See track, 4-person team. Work in progress.

Implementation: my VLM module

One model counts. One explains.

input: FrameRecord (SPEC §6.1)
output: Observation (SPEC §6.2)

Optimization

Each change came from a measurement.

Interaction

The phone reacts to the street.

Map walk-app
Route pathfinding
Observations my vlm
Current location
Seattle Central Library
route v1
Voice check-in my audio-lm + ios-voice
1. A camera sees the sidewalk narrowObservation → A* re-runs → route v2. Router: teammates.
2. The walker leaves the routeVoice check-in. “Help” still asks before calling.

Mock of the UI with example values. Off-path trigger is a button today; 911 is blocked in calling/service.py. Drag the phone to turn it.

1 / 6