Senior ML Engineer

About the RoleWe're hiring a Senior ML Engineer to be the data team's owner of production ML and AI agent systems. You'll take models from prototype to production, build and maintain the low-latency ML API that powers our Next Best Action (NBA) engine, and partner with our HAL team to ship LLM agents that turn NBA recommendations into real conversations with credit union members and partners. This is a builder's role at a builder's moment: NBA is going live, the agent infrastructure is being shaped now, and you'll define how Clutch does production AI for years to come.About the TeamThe Data team today is five people: one data scientist, two data engineers, one data analyst, and one product manager. We're small, ambitious, and shipping fast — two ML models heading to production, an ML API being built, and two AI agents (one customer-facing, one partner-facing) in active development. You'll be the senior technical voice for ML and AI engineering inside the team, and the bridge to HAL, the platform team that builds Clutch's agent runtime. Expect tight feedback loops, real autonomy, and a team that values pragmatism over purity.What You’ll DoWithin 3 months, you will:Take ownership of the ML API that serves NBA recommendations, partnering with the data engineer who's been building it, and harden it for low-latency production trafficShip your first agent tool contract end-to-end: schema design, handler implementation, structured-error contract, unit tests, deployed via HAL's runtimeSet up the eval foundation for our agents: golden transcripts, rubric-based judges, regression suites that run on every prompt or model changeBuild a working relationship with HAL and become the data team's go-to on agent infrastructure decisionsWithin 6 months, you will:Be the primary owner (with data engineer support) of the ML API and the agent tool layer that wraps NBA and our ML modelsHave shipped at least one production-grade agent (customer-facing or partner-facing) with prompt versioning, evals, observability, and multi-tenant gating in placeDefine the data team's playbook for shipping a new ML model as an LLM-callable tool, end-to-endMentor the data engineers on ML/AI patterns so they can confidently support and extend the systems you ownWithin 9 months, you will:Operate as the technical lead within the data team for NBA production AI at Clutch — the person other teams come to when they want to understand how NBA ships ML and agents responsiblyHave measurably improved agent cost and latency (target: 30%+ reduction on P95 latency or per-conversation cost on at least one agent)Be shaping the data team's roadmap for the next generation of ML and AI products, in partnership with the PM and data scientistHelp us decide what to hire next as the team scalesWhat You’ll BringRequired7+ years of engineering experience, with a proven track record of building and shipping production ML systems — you've taken models from prototype to production and own what happens after deployStrong Python — most of the work (ML training, evaluation, the ML API, data pipelines) is in Python, and you're comfortable in production codebases, not just notebooks. Some TypeScript is involved for tool contracts and integration with our agent runtime — you don't need to be an expert, comfort with a second language is enoughTool-design discipline for LLM consumption. Can take an ML model or data source and shape it into an LLM-callable tool with narrow input/output schemas, identity-required and scope-gated dispatch, and structured-error contracts (RATE_LIMITED, UPSTREAM_ERROR, NOT_FOUND) that the agent runtime converts to graceful tool-results instead of crashingEval discipline for non-deterministic systems. You treat evals as the unit-test equivalent for agents: golden transcripts, rubric-based judges, regression suites that run on every prompt or model change. You understand the difference between offline metrics and online evals, and use bothPrompt-shape literacy. You read a system prompt the way another engineer reads code: audience, register, compliance guardrails, template-var allow-list, allowed-tools section. You debug "why did the agent do that?" by reading the prompt and tool descriptions before reaching for model swaps. You've shipped at least one agent where the prompt was version-controlled and reviewed as codeTool implementation rigor. You build handlers behind tool contracts with identity fields read from request context (never from LLM-supplied args), output re-parsed through the tool's schema before return, structured-error throws on every failure path, and unit tests covering both happy path and each named error. You have a story about a tool you shipped, a bug production traffic surfaced, and how you hardened itExperience building and maintaining low-latency production APIs (FastAPI, BentoML, or equivalent), with opinions on serving, batching, and cachingComfortable in AWS (Lambda especially), Docker, and GitHub-based workflowsYo