<aside> 💡

This is a first draft

</aside>

Narrative 1

• LLMS are expected to perform
• Are they a drop-in replacement?
• Messy real world scenario - work that goes into making it useful
  • Good data sets
  • Good feedback loops
  • Knowledge of the model config, parts and the work domain

→ Expectation → Pain point → Tweaks → Resolution

• Audience - Engineering & business - looking to save cost and improve performance.
  • (note)Replacement for expensive SAAS

Type of doc:

• PAN, Indian driver’s license, RC card
• PAN
  • standard format and structure
  • DL and RC layout varies across states and are unstructured.

Existing system

• Takes images of the docs
• Extracts text (OCR): everything.

  • 95% accuracy

  • Even when images are blurry, etc
  • IDFy: blackbox on how it works
  • may or may not be involving an LLM
  • existed way before LLMS
• used for pre-filling forms for onboarding
• Some of it is used for cross-doc verification

Problems with the current system

• Expensive
  • Rs 1500 per 1000 doc