AI agent·In production at Bayer · used on regulatory submissions
Use Summary Table Extractor
A production AI agent at Bayer. It drafts EPA use-summary tables from pesticide labels. Reviewers spot-check before anything goes out.
| Before | Now | |
|---|---|---|
| EPA tables from pesticide labels | 100+ hours by hand | A few minutes |
| What managers do | Write the tables from the PDFs | Spot check, minutes |
| Cost per label | — | $0.50–$0.75 |
| Submission | Waits on transcription | Faster time to market |
Problem
EPA use-summary tables were filled by reading pesticide labels by hand. Crop, rate and use-site data is scattered through the narrative, the rate tables and the appendices. Regulatory managers were spending 100+ hours building those tables from the PDFs.
Value
The tables are now drafted in a few minutes. Regulatory managers only spot-check them, which takes minutes, so the submission gets to market faster. A label costs about $0.50–$0.75 to run.
How I built it
A browser tool with two paths. One uses rules and runs offline, for bulk runs. The other sends the label text through Bayer's AI gateway, then a second pass checks the result. Every row keeps a confidence score, the page number and the source sentence, and you can edit the cell before exporting Excel. One label or many: the 28-column schema stays, and the number of rows follows the labels. A gold-standard check on five labels, 149 pages and 155 rows, came in at 88% field-level accuracy against an expert table.
Deployment
Deployed on Agentrix (container runtime) and through Bayer's AI gateway for model calls. Secrets stay on the server. There is a job queue, a kill switch, a rate limit, a daily cost cap, an audit trail, content-safety flags and drift checks — the Pre-Flight guardrails I apply when shipping agentic tools. It is part of how the department prepares submissions.
Stack
- LLMs (Claude)
- Agentrix
- Bayer AI Gateway
- PDF.js
- SheetJS
- Human-in-the-loop UI