Why this role exists

We're signing enterprises faster than we can model their data.

Every new customer arrives with a data estate nobody has ever mapped. Undocumented tables. Columns named by someone who left in 2019. Business rules that exist only in a sales director's head. Turning that into a semantic model an agent can act on without being wrong is not a pipeline you run. It's a person deciding what correct means in an unfamiliar domain and then proving it.

The role

You'll work on the systems that read messy enterprise data and turn it into a semantic model an agent can actually reason over. Knowledge extraction, natural language understanding, retrieval, and the evaluation that proves any of it is working.

The hard part is not calling a model. It's that "correct" is genuinely difficult to define here. A knowledge graph that looks right and is subtly wrong is worse than no graph at all, because an agent will act on it. Most of the interesting work is figuring out what correctness means for a customer's domain and then proving you hit it.

You'll be close enough to learn from all of them, and the team is small enough that nobody is going to hand you a well-scoped ticket.

If you want a research seat with a publication target, this is the wrong role. If you want clear specs and a defined lane, that's also wrong. You'll be reading unfamiliar customer data, forming your own opinion about what's broken, and shipping the fix.

What you'll do

• Build and improve the NLP and retrieval systems that extract structured knowledge from large, unstructured enterprise data
• Work on the LLM layer: prompting, fine-tuning, RAG architectures, and figuring out which one the problem actually calls for
• Improve the retrieval stack, including semantic search, vector storage, hybrid approaches, and reranking
• Build evaluation. Define what good looks like for a given domain, then build the harness that measures it and catches regressions before customers do
• Keep the pipelines that ingest, process, and serve this data running well at production scale
• Sit with customers and with our product and platform engineers, so what you build solves the problem the business actually has

What success looks like

• First few weeks: You've shipped an improvement to extraction or retrieval quality and told us something we didn't know about where the system is weak
• First few months: You own a real surface of the ML stack. When something regresses there, you catch it before anyone asks
• Beyond: You're the person the team routes a new customer domain to, because you consistently come back with it working

What we're looking for

We care about what you've built, not how long you've been building. Roughly two years of real production experience is the shape this usually takes, but show us the work and we'll judge the work.

• You've put an ML system into production and watched it survive contact with real data
• Strong Python and PyTorch (or TensorFlow). You write code other people can maintain
• You understand transformers, embeddings, and tokenization well enough to reason about them, not just call them
• You've worked with LLMs somewhere real: retrieval, fine-tuning, structured extraction, or evaluation
• You're suspicious of your own metrics. When a number looks good you want to know why before you celebrate
• You learn fast and out loud, and you'd rather ask a blunt question than quietly stay stuck
• You want to be near customers, not shielded from them

Helps, doesn't gate: fine-tuning with LoRA, QLoRA, or adapters; vector databases and hybrid retrieval; knowledge graphs or graph learning; pipelines on Ray, Spark, or Kafka; vLLM; contrastive or self-supervised learning; multi-modal or long-context work; anything data-heavy at enterprise scale.

Why join

• You'll get better here, fast. Four ML engineers and a director who built LinkedIn's knowledge graph. At this team size, that's not a mentorship program, it's just who you sit next to.
• Meaningful equity. You're early. The scope and the upside both reflect that.
• A genuinely unsolved problem. Semantic understanding of enterprise data is the bottleneck on the whole agentic AI thesis, and very few people are working on it at this layer.
• Real enterprise data, right now. Not a benchmark. Production deployments with messy, high-stakes data across insurance, travel, and technology.
• Competitive comp, full benefits (medical, dental, vision, 401k). We sponsor H-1B visas and help with immigration.
• Onsite in San Mateo, flexible hours. We think the hard problems get solved faster in a room together.

Apply

Send us the ML system you're proudest of shipping at [email protected]. Tell us how you knew it was working, and what you'd do differently now.

We value builders over résumés. If this role excites you but you don't check every box, apply anyway and show us the work.