AI
The Problem: Unstructured Prescriptions Drive Precision Manufacturing
ZEISS manufactures precision lenses. Each lens requires structured optical data. The documents that arrive, however, look like anything a doctor’s office might produce: handwritten notes, typed forms, multiple languages, multiple writing systems. None of it is standardized. Wrong data extracted from those documents means a wrong lens ground for a patient. That constraint made automated, accurate extraction a hard requirement, not an optimization.
Why Dapr Agents: Control, Reliability, and Flexibility
Modern models can interpret these documents, but production demanded three properties. First, control: a durable workflow sequences preprocessing, OCR, and LLM calls so the agent cannot produce unpredictable results outside those steps. Second, reliability: state persists after each step, so a completed OCR task does not re-run if something fails midway. Third, flexibility: Dapr’s conversation building block lets engineers point the workflow at a different model by changing configuration alone.
Swapping Models Without Touching Code
Model availability changes fast. If a better or cheaper model appears tomorrow, ZEISS can update a config file and the workflow picks it up. No code change is required. This matters because the workflow itself stays stable while the underlying AI component is replaced. Dapr’s conversation building block is the abstraction that makes this possible, sitting between the durable workflow logic and whichever model the team selects at any point.
Three Rules ZEISS Shipped With
Steinbach summarized the production lessons in three points. First, constrain the AI: a workflow beats letting a model run unchecked. Second, pick the right tool: the system mixes specialized OCR models, general LLMs, and deterministic functions. An LLM should not do what a plain function can handle. Third, decouple model selection from code: config-driven swaps keep the team on the newest models without engineering cycles spent on refactoring.
From Prototype to Production in Two Months
The Dapr Agents architecture took ZEISS from prototype to production in two months. The system required zero labeled training data. Recognition results are on par with a specialized machine learning system built for the same task. Those two numbers, two months and zero training labels, show what a workflow-constrained, config-driven agent setup can do when the architecture keeps the AI bounded and the infrastructure handles state persistence and retry.
Notable Quotes
Wrong data, wrong lens. Fabian Steinbach · ▶ 00:30
don’t use a hammer for a screw. Fabian Steinbach · ▶ 02:04
Don’t use an LLM where a function could do the job. Fabian Steinbach · ▶ 02:12
go from prototype to production in two months with zero label training data required. Fabian Steinbach · ▶ 02:25
Key Takeaways
- Durable workflows constrain AI agents and prevent unpredictable outputs in production.
- State persistence after each step stops expensive OCR tasks from re-running on failure.
- Swapping models via config, not code, kept ZEISS on the best available model without engineering overhead.
About the Speaker(s)
Fabian Steinbach is a Software Architect at ZEISS. Since spring 2020 he has focused on building and running software with Dapr, with particular emphasis on cloud engineering and platform development.