Founding Machine Learning Engineer

Arcade•San Francisco, CA

35d•$230,000•Onsite

About The Position

Arcade is hiring a principal machine learning engineer to build the future of agent capabilities. This role involves applying classic machine learning techniques to enhance AI agents, owning the end-to-end model pipeline from data ingestion to publishing, and obsessing over model evaluation to ensure accuracy and reliability. A key aspect of this role is building enterprise-ready models that can run within customer environments, including air-gapped networks. The models need to be small, fast, accurate, and shippable into a customer's VPC. You will report directly to the Head of Engineering and own the models critical to every agent call served. The first model for agent recommendation is already built and patent-pending, along with its training pipeline, and you will be responsible for its productionization, defining the ML stack, and setting future patterns. This role emphasizes shipping product quickly, with a goal of delivering models and documentation within the same quarter.

Requirements

7+ years of software engineering experience, with 4+ years training and shipping production ML systems. Formal title matters less than the work.
You've trained or fine-tuned models that went to production and moved a metric customers cared about, not just a leaderboard number.
Expertise in production agent systems: harnesses, memory, skills, tool use, and sub-agents. You know the nuances well enough to help decide what we build, and have intuition about what will actually work before we build it.
Know how and when fine-tuning works and how to apply it. What data is needed, and how to derive it from what raw data telemetry gives you.
Evals you'd defend in a review. You have the statistics fluency to say whether a small delta is real or noise.
You've deployed models under real constraints like tight latency budgets, limited GPUs, or someone else's infrastructure (vLLM, ONNX, TensorRT, llama.cpp, or equivalent).
Strong Python for training and ML work, plus TypeScript or Go for the production services that serve your models.
You pick the stack, write the doc, and can still defend the decision a year later.
A do-er, not a researcher-in-residence. You'd rather ship a working v0.5 next week than a polished v2.0 next quarter.
Comfort with ambiguity: early team, a charter that will expand, decisions made with incomplete data.
An insatiable desire to ship.

Nice To Haves

You've shipped ML into enterprise on-prem or regulated environments (financial services, healthcare, government).
Tool-use benchmark or eval work: BFCL, τ-bench, ToolBench, MCP evals, or equivalent.
Familiarity with the MCP (Model Context Protocol) ecosystem. Extra bonus if you've filed an issue against the spec.
You've built training-data pipelines from production traces with privacy controls in place.
You've been the first ML hire somewhere before, and you'd do some things differently this time.
Open-source contributions or published work that survived contact with other engineers.
Experience at an early-stage startup, and you loved it.

Responsibilities

Own the training pipeline: Run Arcade's ML pipeline end to end, from data and training through evaluation and release. Make it reliable enough that shipping a new model is routine, not an event.
Build the models: Fine-tune and train models for tool selection, routing, retrieval, and agent memory. Use whatever gets the job done distillation, embeddings, rerankers and more.
Expand the use cases: Take our model into new territory, starting with search and recommendation over tools and agent context.
Measure what matters: Build the eval system. That means offline evals against real agent traces, online measurement in production, and head-to-head comparisons with Claude, GPT, and Gemini, so every model decision has data behind it.
Ship on-prem: Quantize, optimize, and package models for customer VPCs and air-gapped environments, and work with the Runtime team on how they're served.
Turn telemetry into training data: Build the loop from production agent traces and tool-call data to better models, within the data boundaries our enterprise customers require.
Set the direction: Choose the ML stack, write the strategy, and help decide who we hire next into ML.
Decide what to build: Use your depth in modern agent systems (harnesses, memory, skills) to pick where models make agents meaningfully better, and to call early whether an approach is going to work.
Use AI to compound your own output: Projects that take a week today should take a day next time.