

Giving New Life to Unstructured Data with LLMs and Agents
133 snips Jun 6, 2025
Anant Bhardwaj, Founder and CEO of Instabase, specializes in automating the management of unstructured data. In this engaging discussion, he delves into how large language models (LLMs) are transforming the processing of unstructured documents, enabling innovations like identity verification via WhatsApp. Bhardwaj shares insights on the limitations of traditional robotic process automation and the significance of predictability in AI solutions. He envisions a future where AI agents autonomously handle complex workflows, reshaping enterprise automation.
AI Snips
Chapters
Transcript
Episode notes
Early Challenges in Unstructured Data
- Early unstructured data extraction efforts were brittle and unreliable due to fixed templates and hard-coded rules.
- Introducing layout-awareness with coordinates improved model understanding of documents beyond text sequences alone.
Layout-aware Language Models
- Adding X and Y coordinate encoding into language models revolutionized document understanding.
- This spatially aware approach outperformed prior models by capturing document layouts effectively.
Prioritize AI Predictability Over Perfection
- Enterprises should accept AI systems with predictable error rates rather than demanding perfection.
- Build escalation processes for humans to review uncertain cases and maintain reliability.