Briefly

Local-first legal filing assistant for automated matter classification

Key Metrics: Zero Cloud Telemetry • 4,096 Token Constrained JSON

System Overview

A local-first filing assistant built for a lawyer’s real habit: download now, file later. Watches download folders, extracts text from legal PDFs and DOCXs, and prompts an open-weight model (Gemma) to classify matters with strict confidence gating.

Architectural Decisions & Build Log

Engineered over a weekend to solve confidential file clutter on resource-constrained laptops without leaking private legal matters to public cloud models.

ENGINEERING SPECIFICATIONS

Local Ollama chat API using compact Gemma 4 E2B with memory-safe context limits.

SQLite transaction log ensures duplicate-destination protection and atomic file moves.

Strict confidence gating: files below threshold stay safely in inbox for manual review.

LESSONS LEARNED & TAKEAWAYS

▸Local AI in legal workflows must prioritize privacy over cloud convenience.

▸Scanned PDFs without OCR should fail gracefully rather than hallucinating classifications.

TECHNOLOGIES & RUNTIMES

Python 3.10+Ollama GemmaSQLitePyPDFpython-docxLocal-First