The Fifth Elephant 2025 Annual Conference CfP

You know KYC – Know Your Customer. It involves sifting through and checking numerous documents, such as PAN cards, Driver’s Licenses, and RCs, often in different formats from various Indian states. Sounds like a job for today’s super-smart AI models, LLMs, right? That’s exactly what we thought when faced with the significant cost of our third-party OCR service – approximately Rs 1,500 for every 1,000 documents. This talk is about my exciting experiment to see if LLMs could cut these costs and give us more control.

I’ll share my journey from starting with initial prompts to achieving around 95% accuracy for Driver’s Licenses by iterating through many different “instructions” for the AI and carefully adjusting its settings. I did this so you don’t have to go through the same trials. You’ll get a candid look at the surprising problems I ran into, such as the AI sometimes “skim-reading” and missing details, even in clear images, and the challenges of handling messy or inconsistent document data. The short version: AI wasn’t a magic bullet at first, but we eventually got there. For the more extended version and to see how you can apply these lessons, join my talk!

Key Takeaways

• Learn practical ways to guide LLMs to accurately read text from documents
• Understand why having good data and a clear way to check your AI’s work is crucial for making your own document-reading system successful.

Intended audience

This talk is perfect for Engineers, Product and technical leaders who are exploring or already using AI to read documents, extract information, or want to reduce operational costs and validate the performance.

Bio

Priyanga is a passionate backend engineer, who cares about building useful and usable software. Ask her about her explorations to use AI to improve her engineering workflows.

Draft Slides Deck

https://docs.google.com/presentation/d/1CnQkWJYvOQOnOeTX5qFVDMMwgu4_rMsoG6bbN3ZwBuo/edit?usp=sharing

Mind Map

https://www.notion.so/nilenso-software/Mind-Map-Fifthel-OCR-with-LLMs-Expectation-vs-Reality-2110f0425dae80e7a70fe25953d03337?source=copy_link