Global Response Documents

Privacy set the rules before anything was built. The publications and the inquiries are sensitive, so nothing could be stored: no questions, no answers, no finished documents, and no fine-tuning on any of it. That ruled out most of the usual ways to make a system better over time and pushed the work into the architecture and the prompts instead, which is where the agentic AI design came from: instead of one model learning from data, a set of agents that reason well from the publications in front of them. Speed was the hard stretch. Drafting a full document through one long chain of model calls was too slow to live with. Splitting the work into smaller tasks with less context each, running the expert models in parallel, and letting the writer choose the number of improvement iterations in the interface changed both the waiting time and the quality of the result. The client was genuinely satisfied with what came out at the end. The stack was Python throughout: LangGraph for the agentic flow between the expert models, Langfuse to trace every step of it, all running on AWS. The interface began in Streamlit, which was the right call for prototyping with the writers and, honestly, stayed a little too long: we ended up building more complexity into it than Streamlit is usually asked to carry. Later the whole thing was integrated into another tool and rebuilt in React and TypeScript, which is where an interface like this belongs. Publications arrive as PDFs full of tables and formulas, so getting reliable text out of them mattered as much as anything downstream; Mathpix, Mistral OCR and a few other OCR tools were compared and combined for that. Quality metrics and usage KPIs were kept in DynamoDB and shown with Plotly, so the client could see how the system was doing without asking. We tested with the experts themselves: medical writers used the tool on their own inquiries, gave feedback in the chat as they went, and filled in a form afterwards with both scores and written comments. Those rounds shaped the sections, the feedback loop and the assessment.