We all love the demo phase. You hook up an API, the UI updates, and for a second, the software feels like absolute magic.
I recently hit that phase with a project called Incident Lens AI. It is a forensic video analysis suite I have been building to automate crash reconstruction for insurance and legal teams. The goal is to take raw dashcam or CCTV footage and turn it into a defensible liability report.
To validate the idea quickly, I built a frontend-first proof of concept using React, Vite, and the Gemini 3 Pro SDK. I piped the video frames and audio directly from the browser to the LLM and asked it to act as a forensic expert.
And honestly, it makes for an incredible demo.
You drop a video in, and the system instantly starts reasoning about the crash. It generates liability timelines, cites traffic laws, and outputs structured JSON that drives interactive charts on the dashboard. Building it this way let me iterate on the UI and prove the multimodal concept without writing a single line of backend infrastructure.
But as I transition from pitching a vision to building the actual product, I have to face a hard engineering truth. A cool demo is not a defensible legal tool.
The architecture I used to validate the idea is the exact architecture I now have to dismantle. Here is why.
The Security Issue
First, there is the obvious security issue. Hitting a public LLM API directly from a client application is a complete non-starter when you are dealing with sensitive enterprise data and personally identifiable information. No insurance pilot program will ever approve that.
The Hallucination Trap
But the much bigger issue is the hallucination trap.
My current documentation states that the AI calculates vehicle speed using photogrammetry and motion mechanics. The reality is that LLMs are not physics engines. If you ask an LLM to estimate the speed of a car from a 2D video without precise camera calibration, it is just guessing. It might sound incredibly confident, but in a courtroom setting, an "AI-estimated" speed calculation would be destroyed by opposing counsel in seconds.
You cannot build a forensic tool on prompt engineering alone.
The Hybrid Architecture Pivot
So, I am moving away from the pure LLM wrapper approach and building a hybrid architecture.
I am shifting the heavy lifting to a secure Python backend. The new pipeline will rely on deterministic computer vision models like OpenCV to extract hard, mathematical data from the footage, such as pixel velocities and exact collision coordinates. Once I have those concrete numbers, I will feed them into established physics formulas to get the actual speed and force.
Only then does Gemini re-enter the picture. I will pass those verified, deterministic numbers to the LLM so it can do what it actually excels at: cross-referencing case law, synthesizing the timeline, and writing the final human-readable dossier.
Building in the public safety and forensics space requires an incredibly high bar for trust and accuracy. It is easy to get caught up in the magic of what generative AI can do out of the box.
I am leaving the current repository up as a proof of concept because it perfectly illustrates the vision of where multimodal AI is heading. But the real engineering work of making it secure, deterministic, and legally defensible starts now.
If anyone else is navigating the jump from AI prototype to production in a zero-trust industry, I would love to hear how you are handling it.
You can check out the frontend prototype here:
Sherin-SEF-AI
/
Incident-Lens-AI
Incident Lens AI is a professional-grade forensic video analysis suite. It transforms raw crash footage into a defensible legal case file in seconds.
Incident Lens AI 🔍⚖️
Professional Forensic Video Analysis & Accident Reconstruction Platform
https://youtu.be/QUVeahUrCTg?si=0KiQewMFjjllYqv4
Incident Lens AI is a production-grade application designed for insurance carriers, legal defense teams, and fleet safety managers. It leverages the multimodal capabilities of Google Gemini 3 Pro to transform unstructured video evidence (dashcam, CCTV, bodycam) into legally admissible forensic reconstructions.
Unlike standard video players, Incident Lens AI "reasons" about the footage in real-time, calculating vehicle speeds, inferring traffic signal states from indirect visual cues, and citing specific legal statutes for fault determination.
🚀 Key Features
🧠 Autonomous Reconstruction
-
Physics Engine: Automatically calculates vehicle speed (
$v=d/t$ ) using photogrammetry and motion blur mechanics. - Signal Inference: Deduce the state of occluded traffic lights by analyzing cross-traffic flow and pedestrian behavior.
- Debris Field Analysis: Reverse-engineer impact vectors based on glass shard trajectories and fluid spray patterns.
⚖️ Legal Admissibility
- Search Grounding: Uses the Gemini…
Top comments (0)