All projects
AI & ML · Software Engineering · Cloud & DevOpsJan 2026
VeriJust — AI Audio Forensics (Amazon Nova AI Hackathon)
Built for the Amazon Nova AI Hackathon: a containerised platform that verifies the authenticity of audio and video evidence for project auditing, then runs credibility analysis on the transcript and compiles a forensic report with a chain of custody.
- Amazon Nova AI Hackathon, Jan 2026
- Hackathon
- Amazon Nova AI Hackathon, Jan 2026
- hash on every file, with a recorded chain of custody
- SHA-256
- hash on every file, with a recorded chain of custody
What it does
- Evidence intake: uploaded audio and video files are hashed with SHA-256 and stored in encrypted S3, accessed through presigned URLs
- Signal integrity checks: waveform-level inspection for signs of splicing or re-encoding before any semantic analysis runs
- Credibility analysis: Amazon Nova on Bedrock reviews the transcript and returns a confidence score for each segment
- Tone and sentiment: AWS Comprehend classifies each transcript segment
- Speaker-aware transcripts: diarized, timestamped segments stored with pgvector embeddings for semantic search
- Chain of custody: every action on a piece of evidence is recorded with its actor and timestamp
Architecture
A Next.js frontend talks to a FastAPI backend using JWT cookies. Long-running analysis runs on Celery workers, so uploads return straight away with a task ID to poll. PostgreSQL with pgvector stores cases, transcripts and embeddings. The stack runs in Docker containers, deployed on Amazon ECS with S3 for evidence storage.
Status
Early development: the core pipeline is in place and the remaining services are being wired up.