You'll own the Generation half of Document Intelligence: turning multimodal primitives (keyframes, transcripts, OCR) into schema-valid entities like SOPs, and holding the line on quality so "fast" never becomes "fast and wrong."
Design and iterate recipe prompts — system, few-shot, and context assembly — for each generation recipe
Define per-entity target schemas and domain validators (procedure step-type rules, field caps) that generated output has to pass
Build generation-quality eval datasets and rubrics, offline and online, running on LLMX's eval pipeline, and close the loop when quality regresses
Assemble multimodal context windows from keyframes, transcript, and OCR so each model call has exactly what it needs
Choose the model and token budget per recipe based on quality, cost, and latency tradeoffs
You'll ride on LLMX for model access and on Attachments for ingestion, and hand off validated entities to the domains that own them. You'll report to our Engineering Lead and work closely with the Processing side of Document Intelligence.
LLM observability / cost awareness, or experience with an eval platform
Document, PDF, or video understanding; OCR; retrieval systems
Light fine-tuning experience, or familiarity with the industrial/maintenance domain
Our mission is to keep the physical world running. Factories, fleets, hospitals and campuses stay up because the people who maintain them have tools worth using. That is what we build.
