I built a workbench that turns a deal data room into a reviewable intake report. Each answer has a status and supporting quote; unanswered questions remain visible for the pricing team to resolve.
Before pricing a reinsurance deal, an actuary reviews a roughly 59-field checklist against product specifications, underwriting guidelines, and an RFP. The work includes finding answers, identifying non-standard provisions, and deciding what still needs to be asked. An incomplete answer can matter as much as a completed field.
The workbench ingests the documents and runs agents in a sandboxed workspace. Section workers return structured results that a script merges into a fixed report template. The report contains the checklist, source quotes, severity-ranked flags, and open questions. A chat panel resumes the same session for follow-up review.
Four workflows share one workbench and an archive of runs. The output changes the reviewer’s task from assembling an intake report by hand to checking evidence, exceptions, and open questions.
Fields are marked found, partial, unclear, or missing. Grounding rules require verbatim evidence and instruct the agent to report when the materials are silent. This keeps uncertainty visible instead of turning every field into an apparently confident answer. The pricing actuary reviews the evidence and retains responsibility for the decision.
The same architecture supports memo drafting, rate tables, and experience-data cleaning. For numeric transformations, the agent interprets the input and writes a specification; deterministic code performs calculations and generates files. Faster models draft sections and a stronger model assembles the report.
Python, FastAPI, headless agent CLI, structured JSON, NDJSON streaming, python-docx, pdfplumber, openpyxl, deterministic HTML / Word / Excel renderers, JavaScript.