Hugo Penichou, CTO and Co-founder of SkyDot, a startup specializing in KYC document processing.
During this workshop co-hosted with SkyDot, participants will be challenged to tackle a real-world agent configuration scenario.
Using a synthetic dataset based on actual KYC emails, the objective is to build a library of robust skills and prompts for the structured extraction of key KYC information using LLMs—without fine-tuning. This workshop goes beyond simple prompt engineering, shifting toward Model Thinking through a multi-step workflow (extraction, structuring, and validation).
The session will combine industrial feedback with a methodological approach. SkyDot will present the business challenges and strategies already tested. Participants will then explore various prompt design strategies, experimenting with different phrasing, instruction structures, and output control mechanisms. The goal is to identify relevant entities within a KYC email and standardize requests into JSON files (confirmations, document lists, questionnaires).
The workshop emphasizes an iterative process: rapid prototyping, testing on examples, error analysis, and continuous improvement. Special attention will be paid to the genericity and maintainability of the solutions to ensure scalability.
To conclude, a collective wrap-up session will allow participants to compare approaches, identify emerging best practices, and discuss the current limitations of LLMs in these use cases, paving the way for further industrialization.