The role
- COMP
- $225K - $350K
- EQUITY
- Competitive equity
- LOCATION
- San Francisco
- WORKPLACE
- On-site
- EXPERIENCE
- 8+ years
- VISA
- None, Visa transfers
- STACK
- Python
- INDUSTRY
- AI, Insurance, Software Development, B2B
/roles — ROLE_159
Staff AI Engineer
Fast-growing startup building AI agents that take on manual, document-heavy work for large insurance companies
The company
Venture-backed startup building AI agents for insurance carriers, MGAs, brokers, and reinsurers, automating work like new-business intake, comparing policies, and reviewing underwriting files.
- STAGE
- growth-stage
- FUNDING
- $30M+ raised
- TEAM
- ~40 people
- FOUNDED
- 2024
- BACKING
- VC-backed
JD — the work
About the role
You would join the agent team at a well-funded company automating the manual work that still dominates insurance, and help shape both the product and how the company builds agents. You will not train models: the work is reasoning and orchestration layers built over off-the-shelf LLMs, measured with rigorous evals on real insurance tasks. The company is deeply technical, with engineer founders and engineers who have real influence on product direction. The role is on-site in San Francisco, with meals provided.
What you'll do
- Create evaluation harnesses that track and improve how accurately agents complete actual insurance tasks
- Try new agent techniques such as multi-agent setups and learning from feedback
- Develop MCP servers, tooling, and orchestration that keep production agents dependable
- Ship long-lived agents that process documents, query linked systems, and carry tasks through to completion
- Follow new LLM and agent research and test the best ideas in production
What they're looking for
- 8+ years building ML and AI systems
- Excellent prompting instincts, favoring the simplest approach that works
- Strong Python skills for production agent systems
- Clear, concise communication, both spoken and written
- Drive, grit, and a strong sense of ownership above all
Nice to have
- B2B SaaS background and zero-to-one experience