AI Engineer (Document Intelligence)

About the team
Most AI demos read one clean PDF. Lumin holds millions of real ones — contracts over a hundred pages, scanned documents, tables, and worse — and the product’s AI features only work if that entire corpus becomes structured data: who signed, what it’s worth, when it expires, what can be cancelled. You’ll build and own that extraction pipeline end-to-end.

What your first six months look like
Make extraction work on the hard cases: long contracts where the key terms sit on page 87, scanned documents, tables and line items, more languages.

Build the real-world golden dataset and wire it into a regression gate: every prompt, model, or pipeline change must prove it didn’t make anything worse.

Answer the cost question: what processing the whole corpus takes, which model gives the best quality per dollar, and what it would take for our own models to beat vendor pricing.

Responsibilities

◦Own extraction quality end-to-end: accuracy, latency, and cost across the whole corpus.
◦Shape how documents get read here — which models, which prompting, which checks. The stack isn’t settled yet, and choosing it well is a big part of the role.
◦Build out the eval practice: datasets, regression gates, live quality monitoring, drift detection.
◦Optimize cost: model selection, caching, batching — make quality per dollar visible, because at our scale it decides what ships.
◦Extend what we extract — new fields, nested structures like tables and line items — together with the product teams who turn it into features.
◦Feed the platform’s search and knowledge features with LLM-ready data.

Your skills and experience

◦2+ years as an AI/ML engineer, data engineer, or equivalent with production Python.
◦Hands-on LLM extraction experience: structured outputs, schema design, prompt engineering on real documents, not demos.
◦Eval discipline: datasets, metrics, rigorous measurement — you’ll define and defend what “accurate” means.
◦Production pipeline experience: queues, async processing, monitoring.
◦Comfortable working in English, spoken and written.
◦Self-directed: you can own the pipeline and its quality — you get goals, not step-by-step instructions.

NICE TO HAVE

◦Hands-on vision/OCR experience: VLMs for documents, tools like PaddleOCR, and judgment about when classical CV beats a VLM.
◦Fine-tuning experience or strong interest (LoRA, distillation).
◦RAG experience — chunking, embedding, vector store and retrieval quality.
◦Multilingual NLP; data pipelining.