Evident AI

Automatically transforms bodycam footage into professional evidence reports using AI.

Comment 2

Inspiration

Evidence processing is the bottleneck. Officers spend hours transcribing body-cam footage and writing reports, which is time not spent serving the community. Vendors charge thousands and take days for basic transcripts. I realized there’s a better way: automate the grind while preserving legal standards. Evident AI was born, an AI evidence-processing copilot that turns body-cam video into a professional, review-ready report in minutes, anchored with blockchain proofs.

What it does

Evident AI orchestrates specialized AI agents across the full evidence lifecycle:

  • Full-stack analysis: Extracts audio and transcribes speech with Whisper + error correction and speaker tagging.
  • Visual intelligence: Key-frame analysis (GPT-4V) to surface people, vehicles, objects, and salient actions.
  • Professional reporting: Generates a structured draft with a timeline, entities, action summaries, and narrative*with citations and confidence scores* for quick human review (no legal determinations made by AI).
  • Blockchain integrity: XRPL anchoring + XLS-70 credentials for officer identity, producing cryptographic attestations and a tamper-evident chain of custody.
  • Court-ready output: One-click PDF/Docx with officer checklist, digital signatures, and an audit log (hashes, model versions, timestamps).

TL;DR: Body-cam in → transcript + key frames → cited, structured report → signed & anchored. Minutes, not days.

How we built it

  • Backend: Django 5.2.6 (REST) behind Gunicorn + Nginx
  • AI Agents: OpenAI Agents SDK coordinating transcription, vision, and report-writer roles with structured JSON I/O
  • Video: OpenCV + MoviePy for frame sampling and audio separation
  • Blockchain: XRPL integration; officer XLS-70 creds; memo-based anchors with SHA-256 hashes per stage
  • Database: PostgreSQL (prod), SQLite (dev)
  • Deployment: Dockerized; runs on Heroku/Railway/Vercel
  • Security: Fernet encryption for secrets; audit trails with immutable source hashes

Challenges

  • Agent orchestration: Retries, schema validation, and timeouts across multiple models/services
  • Large files: GB-scale videos; streamed processing and memory-safe chunking
  • Evolving standards: XLS-70 is emerging: DevNet only with limited use case examples
  • Security & compliance: Encryption at rest/in transit, RBAC, and explicit human-in-the-loop review
  • Production hardening: Containerization, DB tuning, back-pressure and job queues

Accomplishments

  • End-to-end demo: Upload → transcript + key frames → cited report → signed PDF with XRPL anchor
  • Human-in-the-loop UX: Confidence chips and inline citations (hover to see the exact timestamp/frame)
  • Integrity by design: Original media hash preserved; every export logs model versions and anchors

What we learned

  • Practical agent hand-offs with the OpenAI Agents SDK (tools, structured outputs, recovery paths)
  • XRPL anchoring patterns and XLS-70 credential flows for attestations
  • Real-world video optimization and keeping AI pipelines responsive
  • Building for policing means: assist, cite, and log

What’s next

  • Pilot with agencies: Run side-by-side with existing RMS workflows to measure minutes saved per report and edit-accept rates
  • Model upgrades: Domain-tuned entity recognizers and policy-aware templates (traffic stop, use-of-force, etc.)
  • Mobile ingest: Secure field upload from camera to case
  • Analytics: Case stats, throughput, and time-savings dashboards
  • RMS adapters: Push approved drafts into Axon/Mark43/Tyler/Niche via export & API connectors

Built With

Share this project:

Updates