This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady
#
What I Built
Satbara Saathi (सातबारा साथी) explains a Maharashtra 7/12 (Satbara) land extract in plain Marathi, Hindi or English.
A 7/12 is dense, jargon-heavy and often scanned badly. Anyone buying or inheriting farmland ends up waiting at the talathi office or paying a middleman to read it to them. The lazy way is to upload a photo and get the answer back:
-
Preprocess the image and check its quality
-
OCR it and extract structured fields
-
Run 11 deterministic red-flag rules (loan/boja, multiple owners, minor owner, tenant, Class 2 land, inheritance mutation, area mismatch, and more)
-
Have an LLM explain the findings in simple language
-
Read the summary aloud (TTS) The most important design decision is what the AI does not do. The rules engine contains no AI. The LLM only explains flags it is handed, and it is told never to call land "safe" or "valid". The tool gives no legal advice and says so.
#
Demo
#
Code
Understand Maharashtra 7/12 (Satbara) extracts in plain Marathi / Hindi / English.
What it does #
-
Reads a 7/12 image (scanned or digital)
-
Extracts structured data via OCR
-
Runs deterministic red-flag rules (no AI in the rules engine)
-
Explains findings in plain language via LLM
-
Speaks the summary aloud (TTS)
What it does NOT do #
-
Give legal advice
-
Decide if land is "safe"
-
Store your images or personal data
-
Process 8A, mutation extracts, or other states (yet)
Quick start #
Architecture #
Privacy #
- Images deleted after processing
- No PII in logs
- Anonymous impact counters only
- Consent gate before upload
License #
Apache-2.0
#
How I Built It
Backend: FastAPI + Pydantic. Providers sit behind small interfaces (OCR, LLM, TTS), so each can be swapped by config. #
Rules:app/rules/mh_7_12.yaml , with bilingual (Marathi/English) titles, explanations and "what to do next" steps per rule. They are plain data, so a revenue-law professional can review them without reading Python. #
LLM: the explain layer targetsSarvam-M , an open-weight 24B model (post-trained on Mistral Small 3.1) with Marathi and Hindi support. A post-check rejects explanations containing names or long numbers that aren't in the source document, and retries once before failing. #
Privacy: consent gate before upload, jobs held in memory only, and the only thing persisted is an anonymous counter DB (no names, no survey numbers, no images). #
Tests: 16 tests pass using fake providers, with CI on GitHub Actions.
Honest status: this is a one-week build. The Sarvam and OCR adapters are written, but the running server currently uses fake providers, and structured extraction from real OCR output is not finished. The rules engine, validation, explanation safeguards, consent flow and UI work end to end on sample data. The rules have not yet been reviewed by a Maharashtra revenue-law professional, which is the first thing needed before any real use.
#
Why Does Open Innovation Matter?
A land record contains names, family relationships and where someone's livelihood sits. That is exactly the data you don't want flowing through a black box.
Self-hostable explanations. Because Sarvam-M has open weights, a gram panchayat, CSC centre or NGO could run the explain step on its own server withvllm serve and keep documents off third-party systems. A closed API would not allow that. #
Auditable rules. The red-flag logic is open YAML, not model behaviour, so lawyers and community members can read, challenge and correct it. #
Built for the language. A model post-trained on Indic languages is what makes Marathi output realistic. Apache-2.0 means anyone can adapt this for other states.
The OCR and TTS adapters currently call hosted APIs. Replacing them with open models is the clearest next step.