samay-rajput/clause-guard

# ClauseGuard AI-assisted Government Resolution (GR) drafting for Maharashtra — VJTI AI Hackathon 2026 ("Shaping Digital Governance", Problem 1: MAHA-GR-ALIGN). **Live demo**: clause-guard.streamlit.app (cloud LLM mode; the four sidebar demo scenarios are pre-cached — fully local Ollama mode is the recording setup, see `DEPLOYMENT.md`). Drafts and reviews Maharashtra GRs against a corpus of real ones: 1. **Cross-departmental conflict detection** — retrieval narrows candidate clause pairs, an LLM (and optionally a fine-tuned MuRIL cross-encoder) screens them, and flagged conflicts get a drafting-officer-grade explanation citing the conflicting GRs. 2. **Reference tracking** — extracts cited GR numbers/codes, links them to the corpus, flags stale (superseded) references, suggests successors. 3. **Template compliance** — deterministic scorecard against Manual of Office Procedure conventions (header, Read block, preamble, numbered clauses, financial-sanction block, budget head, signature formula, code). 4. **Bilingual terminology** — Marathi ↔ English legal/administrative glossary. 5. **Drafting assistance** — a deterministic template scaffold guarantees every mandatory section in both languages; an optional AI layer (local Ollama in ollama mode — the brief never leaves the machine) expands an officer's brief into the preamble and clauses, digit-audited so no numeral is invented or dropped. 6. **Audit trail** — every review and generation appends to a local JSONL: the suggestions offered, the deterministic results, and a unified diff of the officer's edits between submissions. 7. **Export** — the draft and the review report as DOCX (Nirmala UI) or PDF (embedded Noto fonts with harfbuzz shaping, so Marathi conjuncts render correctly). ## Corpus note The live portal gr.maharashtra.gov.in returned 503 throughout the hackathon window (Jul 26–28, 2026). The corpus was therefore built from the **Wayback Machine's archive of the same portal** (≥20,000 archived * …