Read this first. This role is not for you if:

• Your LLM work is limited to ChatGPT, Copilot autocomplete, or a notebook demo that never ran in production

• You've never built a labelled eval set, or you can't say what precision/recall your last LLM check hit

• You've only worked with tables of thousands of rows, not millions or billions

• You've only maintained pipelines someone else designed, and never owned an architecture decision end to end

• Your experience is mostly services or consulting delivery, moving between client projects without long-term ownership of a platform

• You need detailed specs, fixed hours, or a manager telling you what's next

• You don't use AI coding tools daily, or you can't show work you shipped with them

Firmable sells company and people data to sales teams across Asia Pacific and the US, and we're taking it worldwide. The hard problem is keeping billions of records right while companies hire, restructure, move and rebrand every day. More and more of that work runs through models and agents we build into the pipeline.

What you'll work on

You'll design the warehouse and transformation layer that sits under everything we sell, and decide where models belong in it. Problems on the table right now:

• Running LLM extraction and entity resolution over millions of records a day without cost or error rates getting away from us

• Deciding which checks stay as dbt tests and SQL, which go to a cheap decision model, and which need a frontier LLM, then writing that standard down so the team can apply it without you

• Eval sets and regression suites that block a bad prompt or model change before it touches customer data

• Catching drift when a vendor quietly updates a model under us

• Keeping Snowflake fast and affordable at billions of rows

You write the reference implementation and the team builds on it.

Must haves

• 5–7 years building production data platforms

• Strong Python and SQL, plus dbt, Snowflake, Airflow and AWS

Nice to have: Braintrust, Promptfoo or Inspect · Logfire or OpenTelemetry · embeddings for entity resolution · fine-tuning small models · Spark or streaming · B2B data · GDPR/SOC2

How we build

We write code with coding agents like Claude Code by default, and engineers spend their time on design, review and evals. In production we pick the model for the job: frontier LLMs for messy extraction and reasoning, small typed decision models when the answer is a yes/no, a choice or a score, and MCP servers and agent SDKs when a model needs tools. When something new ships, we run it against our evals that week. Every LLM call is logged with its prompt version, model and cost.

Small teams, no layers, weekly releases moving to daily.

The deal

Competitive base and meaningful equity. Evals, drift detection and cost controls for the warehouse are mostly unbuilt, so the design is yours.

Firmable is an equal opportunity employer. We believe diverse teams build better products.