ClearDoc

People sign rental agreements, employment contracts, and NDAs without understanding the dense legal language inside them — and risky clauses go unnoticed until it's too late.

ClearDoc extracts text from scanned or native PDFs via OCR, auto-detects the document type, runs it through a modular rule registry that flags clause-specific risks, and uses Gemini to summarize the document and answer follow-up questions in plain English and Tamil — before the person signs anything.

• OCR + PDF pipeline extracts text from scanned and native documents
• Modular rule registry — new document types (rental, employment, NDA) plug in without touching core logic
• Gemini-powered summarization, clause risk-flagging, and click-to-explain chat

• app.py — Streamlit UI: upload, results display, click-to-explain chat
• pipeline.py — orchestrates the flow: detect document type → extract facts → run rules → summarize
• doc_types.py — the modular rule registry, one block per document type (rental, employment, NDA)
• extraction.py — PDF/OCR text extraction, kept in-memory only
• llm_client.py — the single point where the app talks to Gemini, so the rest of the app stays provider-agnostic

Team project built for a campus hackathon — worked on the extraction pipeline and the plug-in style rule-registry design that lets a new document type (rental, employment, NDA, etc.) be added without touching the core pipeline logic.

• Extraction accuracy depends on document clarity — heavily degraded scans can confuse auto-detection and need manual document-type selection as a fallback.
• The rule sets encode specific, known checks — they're not a substitute for legal advice, so the app says so visibly in the UI rather than overstating what it can catch.

• Designing a rule registry as a plug-in system, so new document types are additive instead of requiring changes to shared pipeline code.
• Isolating every LLM call behind one module (llm_client.py) so the rest of the app doesn't care which model is behind it.