Salary range - $275k - $325k | Equity - 0.25% | In-person NYC
About Datalab
Datalab trains models that read documents reliably at scale. The world's most important information is trapped in PDFs, scans, and files that can't easily be parsed, and getting it out correctly matters. From frontier AI labs processing training data to Fortune 500s like Siemens extracting decades of engineering records, Datalab is where businesses turn to when extraction has to be right.
We’re at high 7-figure ARR. Anthropic, Legora, Booz Allen, and ElevenLabs are customers. And we have hundreds more across FAANG, frontier AI labs, healthcare, finance, government, and legal. Our tools, chandra, surya, marker, and lift, have 70,000+ GitHub stars and broad developer mindshare. We're backed by founding members of OpenAI, FAIR, and Hugging Face.
Role Overview
We're looking for a Research Engineer to own problems end to end across our models, inference service, and product. You won't just train a model and hand it off. You'll take it from training through benchmarking, into our inference stack, and work with the team to integrate it into our products.
We're a small team that has shipped the current state of the art OCR model, Chandra. Our models collectively have 70k+ Github stars. Our tools are used internally at frontier AI labs like Anthropic, and Fortune 500 enterprises like Siemens.
Our team focuses on training small, efficient models that outperform much larger LLMs on domain-specific tasks (like OCR, structured extraction, tables). We move fast, prioritize practical results, and build tools that are open, reproducible, and built to last. You'll test hypotheses quickly, iterate on results, and balance experimental rigor with shipping to customers.
Day to day:
A typical project might look like: identify a gap in extraction quality on long documents, train and benchmark a new model, optimize it for inference, and work with the team to ship it to users. Concretely:
Train and evaluate models: Train task-specific models (OCR, layout, text recognition, extraction). Explore architectures and training strategies to optimize task performance. This includes our open source models, like Marker, Surya, and Chandra.
Optimize inference: Profile and accelerate model inference across different hardware setups (H100s, B200s, L40s, CPUs).
Ship to product: Work with the team to integrate models into our API and product, helping define how new capabilities surface for end users. You will be involved from model training through integration, although your work will be weighted much more towards the model side than the product side.
Create and maintain datasets: Source, design, and clean datasets for supervised and synthetic training; create reproducible pipelines for data versioning and evaluation.
Experiment and benchmark: Run ablations, track metrics, and publish findings that inform model design and internal research direction.
Engage with users and partners: Occasionally join calls or Slack threads to better understand customer needs and inform your work.
Ideal Candidate
You've shipped models that made it into production. You understand how to balance exploration with delivery, and how to turn research insights into products people actually use.
3+ years experience training, fine-tuning, and evaluating deep learning models
Trained at least one production-grade model or system used in real-world applications
Deep expertise in PyTorch and Python, with strong fundamentals in deep learning (optimization, evaluation, architecture design)
Comfortable with data engineering, benchmarking, and performance profiling across hardware setups
Comfortable with an early stage startup - balance running ablations/benchmarks with shipping velocity
Bonus points if you:
Have experience with OCR, document AI, or structured extraction
Have published work, whether that's a paper, a benchmark report, or a deep technical blog post
Have been a major contributor to open-source projects, especially in ML, vision, or NLP
Enjoy writing about your work and sharing learnings with the community
Interview process
A 30-minute video call to evaluate fit
90-minute live architecture discussion
Culture fit interview/team meeting
At this stage of the company, every interview is somewhat custom, so these phases may be rearranged slightly.
Equal opportunity
Datalab is an equal opportunity employer. We do not discriminate on the basis of any characteristic protected by federal, state, or local law.
If you need an accommodation to participate in our hiring process, or if you believe you have experienced discrimination or harassment at any point in it, contact operations@datalab.to.