Verify Indian land & legal documents, and answer statutory-compliance questions, entirely on your own infrastructure. Confidential deeds never leave the building.
Property due-diligence and statutory compliance in India are still largely manual: a human reads sale deeds, encumbrance certificates, RTC / land records, and cross-checks them against state registries and a stack of Acts. It's slow, expensive, and the one thing you can't do is upload a client's title documents to a third-party cloud OCR. Flaiw does this work locally — extraction, verification, and compliance reasoning with no outbound network by default.
Flaiw is two complementary systems:
A batched, local-only pipeline that turns raw legal documents (sale deeds, RTCs, encumbrance certificates, partition deeds, etc.) into structured, verified results.
Raw documents → object store → OCR → ML field extraction → registry verification → human review → verified result
- Local-by-default, egress-guarded — a dedicated
egress_guard/privacy_guardblocks outbound network during processing, so documents stay on-prem. Optional commercial-OCR fallback is opt-in, not the default. - Hybrid OCR — local OCR first (multilingual: English, Hindi, Marathi, Tamil, Gujarati), with selective fallback only where you allow it.
- ML field extraction — layout-aware extraction (Donut / LayoutLM-family) for the fields that matter in title work.
- Registry verification — pluggable, per-state government-registry connectors with caching (
verification/registry_guard,auto_verify). - Human-in-the-loop — Label Studio annotation + uncertainty-sampled active learning so accuracy improves on your own corpus.
- Built to scale — batched ingestion, Airflow orchestration, Postgres for metadata/results, object-store backends (local / MinIO / S3), designed for large multi-terabyte archives.
A small local utility that answers statutory-compliance questions for a scenario:
- Asks a local LLM (via Ollama) which Acts / Rules / Regulations apply to your scenario.
- Reads the relevant Act PDFs you keep alongside it (ships with Contract Act, Factories Act, Indian Partnership Act, Stamp Duty Act).
- Extracts the concrete documents / registrations / licenses / IDs required, with citations where possible.
- Saves a consolidated checklist to
outputs/.
python app.py --query "I want to purchase a commercial building complex. What documents are required?"Title documents, land records, and client agreements are confidential and often legally sensitive. The managed OCR / document-AI APIs are convenient but mean shipping those documents to a third party. Flaiw is built the other way around: the documents stay where they are, and the intelligence comes to them. That's the whole point for law firms, lenders doing title due-diligence, and corporate compliance teams.
Flaiw is the legal/compliance vertical built on the same on-prem document-AI thesis as my other work — if you need the raw extraction primitives, see OpenExtract (drop-in, self-hosted document extraction) and parakh (accuracy evals + correction loop). Flaiw is the domain layer that turns those primitives into verified title and compliance outputs.
Flaiw is an active, modular system: ingestion, OCR, extraction, registry verification, security/egress controls, and the compliance assistant are implemented as working components, with sample runs in law_firm/data/processed/. It is not a turnkey SaaS — it's a framework you deploy and tune to your state's registries and document mix. Benchmark it on your own corpus before relying on it; accuracy on legal documents is corpus-specific by nature.
On-prem legal/document AI, title due-diligence at scale, or air-gapped deployment for regulated work — tanupam760@gmail.com.