Validation, not prevention, for non-deterministic systems

Speaker at the ING & DataStax AI event in Amsterdam, sharing the stage with Colin Beales (GitHub) and Jeffrey Opdam (ING). The thesis: the probabilistic nature of LLMs is a feature, not a bug — and what makes them production-ready is validation, not constraint.

The talk worked from a small-business case: a Dutch restaurant owner using AI to automate the invoices, contracts, and emails nobody wants to touch. Live demo on a set of highly unstructured (and intentionally erroneous) documents, with backwards compatibility into existing software systems as the design constraint.

Stack: OpenAI for generation, Pydantic for typed contracts, Instructor (Jason Liu) for structured output, Unstract LLMWhisperer for ingest. The same pattern the training reaches for: validation surface as the operating layer that makes non-deterministic systems shippable.