EasyHunt
Title due diligence for home loans. Tech lead at Encegen AI Labs, team of three
Python, FastAPI, Tesseract OCR, Gemini, Supabase
What
Bank lawyers in India check a property's title by reading scanned land records. EasyHunt reads them for the lawyer and drafts the report. It takes up to 20 documents per case (7/12 extracts, mutation records, sale deeds), in several languages and some handwritten, and produces a scrutiny report in the bank's own template for the lawyer to review and approve.
How
Tesseract OCR runs first, with a vision-language model (Gemini) as the fallback when confidence is low. LLM extraction runs at temperature 0, followed by chain-of-title checks and risk flags. Every document is searchable in the original language and in English, and the report is filled from bank-specific templates.
The hard part
Deciding what to trust. Model confidence isn't reliable on handwriting, so every handwritten document goes to human review, however sure the model is. On a legal document, slower beats confidently wrong.
Result
I own the LLM pipeline. 164 automated tests.
What I'd do next
Fine-tune per document type once there's enough volume for each one.
Next projectVoice agent platform