AI research engineer

About the company

This company builds the document intelligence layer for enterprise workflows. The open-source framework is downloaded more than 4 million times per month, and the cloud product has processed more than 200 million documents. Customers include Fortune 500 teams in financial services and consulting, where document accuracy, extraction quality, and workflow reliability are not optional. The company was founded in 2022, has a team of about 40 people, has grown from $0 to nearly $10M ARR in 16 months, and is backed by Greylock and Norwest. The open-source product and enterprise cloud product create a tight loop between developer adoption and production document workloads, which makes model improvements measurable very quickly.

About the role

This is a hands-on applied research role on the core document understanding team. You will own the loop from customer documents to data curation, benchmarks, model training, post-training, and product quality. The work sits between research and engineering: you are expected to make model decisions that improve accuracy, reduce inference cost, and hold up under enterprise documents that are messy, domain-specific, and expensive to get wrong. You will spend time with customers, but it is a small slice of the job: less than 15%. The main surface area is model, data, and evaluation work. This is not a prompt-only role. The benchmark is part of the product contract.

The technical problem

Document AI fails in narrow, expensive ways. Scanned PDFs, tables, long forms, multi-column layouts, domain-specific language, and inconsistent customer inputs all create failure modes that benchmark-only work can miss. The hard part is not training a model in isolation. The hard part is building a system where data quality, benchmark design, post-training, and agent behavior line up with production outcomes and cost constraints. Every improvement has to move a three-way objective: quality, latency, and cost.

What you'll own

• Document-focused model training: develop and train vision-language models for document processing, including data curation, synthetic data generation, and training runs that improve real extraction quality.
• Benchmark design: build and maintain evaluation sets that reflect real customer failure modes, not just convenient offline metrics.
• Base model evaluation: compare candidate models, understand where they break, and choose the smallest effective model or post-training path for the task.
• Post-training: adapt models to specific document tasks and performance targets while keeping serving cost and latency under control.
• Agent quality: improve the agent layer on top of base models so extracted document data becomes reliable downstream behavior, not just better text output.
• Experimentation velocity: build the tooling and process that let the team run cleaner experiments, diagnose regressions quickly, and ship model improvements faster.
• Customer translation: work directly with customers when needed to turn messy product requirements into measurable benchmarks and model capabilities.

Who this is for

You are likely a strong fit if you have:

• 3–7 years of experience in machine learning engineering or applied research.
• Built and shipped model systems end to end, not just run experiments in notebooks.
• Hands-on experience with dataset construction, labeling strategy, synthetic data, or benchmark design.
• Strong error analysis instincts: you can tell whether a model missed because of data quality, evaluation design, or model capacity.
• Experience turning paper ideas into production constraints without losing the core insight.
• Comfort working on multimodal or document-heavy problems, especially where layout, OCR noise, or long-context behavior matter.
• Good judgment about when to fine-tune, when to post-train, and when to leave a base model alone.
• The ability to explain tradeoffs in accuracy, latency, and cost without hand-waving.
• The ability to work directly with product and customers without turning the role into account management.
• You should be able to move from a paper to a benchmark to a production decision without losing rigor.

Why now

The company already has the kind of usage that makes research concrete: millions of open-source downloads, hundreds of millions of documents processed, and enterprise customers who care about correctness at scale. The next step is not more generic model experimentation. It is building a tighter document understanding loop that improves quality, lowers cost, and creates a clear technical edge in a crowded market. This role is how that loop gets built.

This role is not for you if

• You want a pure research role with no production accountability.
• You only want to work on prompting or orchestration.
• You do not want to own benchmarks, data quality, or model evaluation.
• You prefer well-specified tasks over ambiguous, high-leverage problems.
• You need remote-first work or require new H1B sponsorship.

Compensation and logistics

• Base salary: $180K–$250K
• Equity: 0.1%–0.4%
• Location: San Francisco, CA
• Work model: hybrid, with Monday / Wednesday / Friday in office
• Employment: full-time
• Visa: no new H1B sponsorships; TN available; H1B transfers may be considered for exceptional candidates