WorkFER drafting platform
FER drafting platform
A 14-node LangGraph pipeline with 4 ReAct agents on a self-hosted model, hardened for the October 2026 release.
~30%down to<1%
structured-output failures
The numbers
14
nodes in the LangGraph pipeline
4
ReAct agents
1,769
tests passing for the October 2026 release
Notes
The FER drafting platform is a team project: a 14-node LangGraph pipeline with 4 ReAct agents, running on a self-hosted model.
My part
I hardened the pipeline for the October 2026 release, which shipped with 1,769 tests passing. I also cut structured-output failures from about 30% to under 1%.
What the system does
It is a five-service system. It takes an Indian First Examination Report and Form 3 and produces a drafted response, objection by objection.
Getting the documents
Documents are acquired from InPASS, USPTO Global Dossier, EPO, JPO, KIPO and WIPO by Selenium workers behind a job queue with retry and orphan recovery. Discovery is split from download, so attorneys pick the 10 to 20 relevant documents out of 200+ per patent family.
Reading them
OCR (PaddleOCR and Azure) feeds a content-based LLM classifier that sorts documents into 10 types using head and tail excerpts, because Indian specifications put the claims at the end. Form 3 extraction includes a row-count self-check.
Drafting in two phases
A 10-state FER lifecycle in the orchestrator (Prisma and Postgres) runs the drafting in two phases: strategy and amended claims first, then attorney confirmation, then the final response. A queue and boot-time recovery protect drafts in flight. The React workspace has a stepper, document picker, manifest editor, a chat against the drafting engine, and DOCX and PDF export.