Machine Learning Engineer

Shepherd•San Francisco, CA

About The Position

Shepherd is a technology-driven Managing General Underwriter (MGU) transforming commercial Property & Casualty insurance for high-hazard industries. Our mission is to make risk frictionless for the builders and operators shaping the physical world — protecting progress from concept through construction and into decades of operation. We’re building the fastest, smartest commercial risk platform, where underwriting expertise, data, and automation work together to deliver: Faster decisions Smarter, more accurate pricing Better risk outcomes With Shepherd, safety, speed, and quality no longer trade off against one another — they compound. We’re not just modernizing insurance products. We’re building the risk infrastructure for the next generation of financial services, where technology, underwriting, and partnerships operate in harmony to support the world’s most important industries — and the progress they make possible. We think about underwriting autonomy the same way Waymo thinks about self-driving cars. Not as a binary switch, but as a graduated progression through defined capability levels. Today, Shepherd sits at the border of L1 for our first Operational Design Domain. You will build the ML systems that carry us from L1 to L3 and beyond. Every model you ship, every feedback loop you close, and every confidence threshold you calibrate is one more autonomous mile driven. You will be Shepherd’s first Machine Learning Engineer, embedded in the Fully Autonomous Underwriting (FAU) team. This is a high-ownership, high-ambiguity role. There is no existing ML platform to inherit, no established model registry to maintain. You will build those things. You have the opportunity to define the ML function from the ground up at a company building something genuinely new in a large, underserved market You will work directly with underwriters to deeply understand the domain, and translate that understanding into ML systems that get meaningfully better over time. You will own the full ML lifecycle – from data through to production – and be the connective tissue between the domain expertise that exists in the business and the systems we’re building to scale it.

Requirements

4+ years of industry experience building and shipping ML systems end-to-end, from raw data to production models
4+ years of industry experience building and shipping ML systems end-to-end, from raw data to production models, including experience with model deployment platforms (e.g., AWS Sagemaker)
Experience finetuning SLMs/LLMs, with a preference for experience using techniques like RLHF, DPO, or LoRA.
Deep proficiency in Python and modern ML frameworks (PyTorch, HuggingFace, Tensorflow, OpenAI Gym/Gymnasium or similar)
Experience with LLMs in production: prompt engineering, structured outputs, tool use, evaluation, and cost/latency tradeoffs
Experience building reliable models with limited labeled data, including synthetic data generation, data augmentation, or similar techniques
Strong evaluation instincts: you know how to define what ‘better’ means before you build, not after
Comfort with ambiguity, highly autonomous, and a bias toward building something real over architecting something perfect
Excellent collaboration skills. You will spend significant time with non-technical underwriters and need to earn their trust

Nice To Haves

Familiarity with document parsing, information extraction, or NLP on unstructured business documents
Background in insurance, finance, or other high-stakes structured domains where model errors have real consequences
Experience with agentic frameworks or multi-step LLM orchestration (LangChain, LangGraph, or custom)
Confidence calibration experience: isotonic regression, Platt scaling, or similar techniques
TypeScript proficiency. Our platform is TypeScript-heavy and cross-functional contribution is valued
Familiarity with data pipelines: SQL, dbt, Spark, or equivalent
MS or PhD in a quantitative field (ML/AI, Statistics, Math, Physics)

Responsibilities

Design, build, and ship ML systems that power autonomous underwriting decisions in production
Build and close the feedback loops that turn human underwriter behavior into training signal and compounding model improvement
Develop confidence scoring and evaluation frameworks that define when the system is ready to take on more autonomy and when to step back
Work with large language models to build reliable, auditable, and improvable agentic workflows across the underwriting lifecycle
Partner directly with underwriters to extract domain knowledge, validate outputs, and earn the trust required to expand the system’s operating domain
Contribute to the observability, monitoring, and guardrail infrastructure that keeps AI underwriting safe as autonomy scales