tokens&
For enterprises
tokens&

Find tools, check provider offers, save a build plan, and share your work when you’re ready.

For buildersFor enterprises

For builders

  • Startup credits and perks
  • Agent Skills
  • Publish a project

For enterprises

  • Start free company workspace
  • Submit a tool, product, or perk

Community

  • Community
  • Newsletter
  • Events
Xin

© 2026 tokensand, LLC. All rights reserved.

  • Terms
  • Privacy
  • Security
  • Data Processing
  • Status
  1. Hackathon
  2. Project gallery
  3. CloseCall
Anonymous builderabout 1 hour agoJudging locked: Event build

CloseCall

A flight recorder for intersections. CloseCall turns traffic-camera video into a ledger of near-misses, with the clips, a plain-language search and a fix memo, so a city can make the case for a redesign before the crash instead of after it.

Review the project

Start with the source code, then open the demo or video if available.

View GitHub repository
Visit project websiteProject gallery
Project description
For city traffic engineers. Today an intersection only "counts" as dangerous once someone is hurt, because crash reports are the only evidence a council accepts. Meanwhile, the cameras already record every close call, and nobody watches them. What CloseCall does: Track: Ultralytics YOLO26 tracks every road user. Post-encroachment time (PET) under 4 s flags a conflict, and a physics filter keeps only real crossing paths. Detected heads and plates are pixelated before any model sees a frame. Triage: each candidate becomes an 8-frame storyboard, and three open models (Qwen, MiniMax, Gemma) on W&B Inference on CoreWeave vote; two of three must agree. Every vote is traced in W&B Weave. Search: NVIDIA Nemotron embeddings let the engineer search in plain words ("turning car cutting across a pedestrian"). Identity questions are refused. Act: NVIDIA Nemotron-3.5 Lightning drafts a fix memo in which every claim cites a clip id, and a council-packet CSV export (built with Cursor) goes to the meeting. Play: "Spot the Near-Miss", a 10-round game: you vs the 3-model vote vs the label. Measured, not claimed: from 141 s of footage, 1,165 raw flags → 51 candidates → 12 clips to review. The vote caught 3 of 4 labelled near-misses across all 51 events, and 1 of 2 on held-out events (blind, pre-registered labels; every step is in the repo). A VAST VSS client is built behind a storage adapter, ready to connect.
Project links
  • GitHub repository
  • Project website
Tools used
  • NNVIDIA
  • C(CoreWeave (Weights & Biases)
  • S(SpaceXAI (Cursor)
  • VDVAST Data
  • UYUltralytics YOLO26
Tools used
  • NNVIDIA
  • C(CoreWeave (Weights & Biases)
  • S(SpaceXAI (Cursor)
  • VDVAST Data
  • UYUltralytics YOLO26