Sightline
When a car crash ends up in court, a lot of it comes down to a simple question: could that witness actually see what they say they saw?
Core trick
Argon writes the 3D crash scene as code, and deterministic line-of-sight checks test what each witness could see from their stated position.
Description
When a car crash ends up in court, a lot of it comes down to a simple question: could that witness actually see what they say they saw? Right now, answering that means hiring a reconstruction expert, waiting weeks, and paying someone to sit through hours of deposition video.
Sightline does that first pass in minutes. You upload the case file (police and NTSB reports, exhibits, videotaped depositions) and Gemini 4 Argon reads and watches all of it in one go. Our demo uses the real NTSB docket for a 2016 crash in Florida: 212 pages plus almost three hours of (simulated) deposition video, about 849k tokens. Argon pulls out the vehicles, the timing and where every witness stood, with citations, and a lawyer checks and confirms it. Then Argon writes the entire 3D scene as code in a single response. Plain geometry and physics checks with no AI in them test what each witness could really see, and Argon fixes its own world when the checks or the evidence disagree.
The moment we’re proudest of: one witness swears nothing blocked her view, but she never says where she was standing. Nearly two hours into her deposition, she just draws an X on a map and says “Right about here.” Argon caught that, measured it, and from that exact spot the bench shelter and a turning truck block her view of the car. That holds from all 25 positions we test around her mark, and 23 independent agents reproduced the result.
You can also type in a brand-new witness statement and test it in seconds, then export a cited report foor the expert. It’s a first draft for a professional, not evidence. The crash facts are real; the witnesses are fictional.
Welddevelopment/Sightline---argon-4-hack-sf★ 00 forksJavaScriptCreated Oct 25, 2026 · last push Oct 26, 2026
README excerpt
# Sightline **Test witness testimony inside a 3D reconstruction of the crash.** Sightline is for personal-injury and insurance-defense lawyers and the reconstruction experts they work with. You upload a crash case: reports, statements, photos, diagrams, and video such as videotaped depositions or dashcam. Gemini 4 Argon reads and watches all of it in one call and extracts vehicles, occluders, timing and witness positions. A human confirms them. Argon then writes a 3D world of the crash as code, and the loop checks its own work: 1. We render the world from each witness's eyes. 2. Independent geometry and physics checks run, with no AI in them. 3. Argon watches the renders next to the evidence and patches the world. The output is a reconstruction you can stand inside, plus findings that cite report pages and video timestamps, exportable as a PDF report. Built at the Gemini 4 Argon SF Hackathon, Oct 25, 2026. **Track 2: More Dimensions.** - Demo video (1 min): `TODO` - Team: `TODO name (@github)`, `TODO name (@github)` ## Compared with what lawyers pay for today Each pair puts a frame from a commercial forensic animation (left) next to the same kind of view from Sightline (right). Our frames come from the World B reconstruction: Argon wrote it in one pass from 849k tokens of case files, then corrected it once. Each Sightline frame also shows live speeds, the time to impact, and a check of whether the witness could see the car, with the testimony cited. | | | |---|---| |