IncidentWeave

IncidentWeave is an AI-assisted multimodal incident and case investigation platform that helps investigators turn fragmented evidence into a clear, traceable incident reconstruction.

What it does

IncidentWeave brings together evidence such as:

• Images and screenshots
• PDFs and witness statements
• Call logs and messages
• CCTV frames
• Notes and documents
• Audio and transcripts

It then correlates these sources to help reconstruct a chronological timeline, surface contradictions, identify missing or unknown evidence, and support human review.

Fragmented evidence
        ↓
Ingest / extract
        ↓
Cross-modal correlation
        ↓
Timeline reconstruction
        ↓
Contradiction detection
        ↓
Unknown / missing evidence
        ↓
Human investigator review

Core principle

AI surfaces patterns. Investigators make decisions.

IncidentWeave is designed as a human-in-the-loop system. AI findings should remain traceable to source evidence and clearly distinguish observed information, inference, and uncertainty.

Planned technology stack

• React + TypeScript + Vite
• Tailwind CSS + Motion
• Supabase Auth
• Supabase PostgreSQL
• Supabase Storage
• Supabase Row Level Security
• Supabase Edge Functions / server-side application logic
• Google Gemini API for multimodal reasoning
• GitHub for version control

Target workflow

1. Investigator creates an investigation.
2. Evidence is uploaded and securely stored.
3. Evidence is prepared for analysis.
4. Gemini analyzes the evidence across modalities.
5. Structured findings are persisted.
6. IncidentWeave presents the timeline, contradictions, unknowns, confidence and source relationships.
7. The investigator reviews and makes the final decision.

Analysis reliability

The analyze-evidence Edge Function sends actual evidence content to Gemini and treats HTTP 429, 500, 502, 503, and 504 responses as transient. Each request is retried with exponential backoff and jitter, has a 90-second timeout, and can fall back from GEMINI_MODEL to GEMINI_FALLBACK_MODEL (default gemini-3.8-flash). Configure both server-side secrets when deploying the function. A persistent analysis run is marked failed only after the configured models and retries are exhausted.

Security principles

• Keep Gemini API credentials server-side.
• Never commit secrets.
• Keep sensitive evidence in private storage.
• Enforce access through Supabase Row Level Security.
• Do not treat AI output as an autonomous legal or criminal verdict.

Hackathon

HackDays Solan 2026 — Enkindle Club

Theme: Google Gemini API

Category: Software / AI / Open Innovation

License / demo data

Any future demo evidence should be synthetic or openly licensed and clearly identified as such.

Local configuration and deployment

Copy .env.example to .env and set VITE_SUPABASE_URL and VITE_SUPABASE_ANON_KEY. Apply the SQL migrations in supabase/migrations to your Supabase project, deploy supabase/functions/analyze-evidence, and set its server-side secrets: GEMINI_API_KEY, GEMINI_MODEL (optional), and GEMINI_FALLBACK_MODEL (optional). Do not put a Gemini key in a VITE_* variable or commit it to the repository.

The browser uploads evidence to the private evidence bucket, then invokes the Edge Function using the authenticated user's JWT. The function checks the investigation and analysis-run owner before using its service role to download evidence and call Gemini. If Gemini or a dependency fails, the analysis run is persisted as failed and no replacement or mock findings are generated.