Cytora

RiskGPT

In 2023, Juan helped Cytora with their goal of reducing the underwriting process from 2 weeks to 20 minutes using the latest developments in document understanding and language models.

I really want to thank Juan for the contributions he’s made to Cytora over the last year, the work we have done on LLMs has driven a compelling new direction for our product.

— Sam Lewis, VP Product, Cytora

Detail

• fine-tuning smaller (3-7b parameters) open-source models with applications in structured data extraction (fields, addresses), and singular field extraction (“single field experts”)
  • applied: dolly
  • research: llama(+derivatives)/mistral/mixtral, mixtures of experts
• prompt engineering
  • applied: zero/few-shot, chain-of-thought
  • research: prompt chaining, automatic reasoning, auto prompts, function calling
• fine-tuning large foundational models
  • applied: Google’s PaLM2/Vertex, Databricks
  • research: Amazon Bedrock
• research into layout detection (LayoutXLM trained with DocLayNet/PubLayNet)