The Core Challenge: Discovery Overload in Public Interest Law
In civil rights litigation against state agencies and law enforcement, discovery production often results in massive “document and media dumps”—hundreds of hours of body-worn camera (BWC) footage, thousands of scanned PDFs, and messy email archives.
Legal teams face two critical constraints:
- Severe Resource Asymmetry: Non-profit legal investigators must manually review hundreds of hours of raw footage and dense paper scans to identify key infractions.
- Strict OpSec & Protective Orders: Case evidence is bound by court orders and bar confidentiality rules. Uploading sensitive investigative materials to hosted commercial APIs (e.g., OpenAI, Anthropic) violates legal ethics and protective covenants.
Technical Approach: Local-First Systems Engineering
To solve both constraints, I built an end-to-end local document and media processing engine that runs entirely on isolated, in-house hardware.
[Raw BWC Footage / PDF Scans]
│
├───► Audio Extraction ───► Local Whisper ────┐
│ ▼
└───► Image Preprocessing ──► Local OCR ──► [Normalized Text Stream]
│
▼
[Local Embedding & Index]
│
▼
[Investigator Triage UI]
1. Ingestion & Multimodal Extraction
- Optical Character Recognition (OCR): Preprocessing scanned court filings, handwritten field notes, and low-contrast paper documents using tuned binarization before text extraction.
- Audio & Speech-to-Text: Local Whisper models transcribing BWC audio in parallel, producing timestamp-aligned transcripts indexed by speaker turn and incident time.
2. Local-First Indexing & Semantic Triage
- Text streams are chunked and embedded with local transformer models, allowing attorneys to search across multimodal archives using semantic concepts rather than just fragile keyword matches.
- All embeddings and vector indexes remain strictly local with zero internet transmission.
3. Workflow Orchestration & Version Control
- Metaflow manages deterministic DAG pipelines from raw media ingest to feature storage.
- MLflow & DVC track model versions, parameter configs, and data hashes to ensure every step is auditable for court evidentiary standards.
Impact
- Drastic Triage Acceleration: Accelerated investigative document review cycles, allowing legal teams to pinpoint relevant timestamps and key evidence across massive productions.
- Full Data Sovereignty: Maintained strict compliance with protective court orders by keeping 100% of data processing within trusted physical boundaries.