The role

Build the AI workforce. You are the architect of our operational automation platform, not a traditional application engineer. Our agents quote, bind, draft client email, answer the phone, and drive carrier portals, and the runs are not chat turns: a submission sits with an underwriter for three days, a carrier portal times out mid-flow, a reply lands at 6am, and a human has to approve a bind.

What you will own

• The agent systems: carrier browser agents, and the email, voice, bond, and licensing agents

• The retrieval stack: vector search plus a knowledge graph of companies, contacts, policies, and carriers

• The document and extraction pipeline, and the worker fleet behind it

• Carrier and bank integrations, and the client portal

What we are looking for

• 3 to 8 years, with production LLM and agent systems behind you, not demos

• Durable async execution. This is the core of the job: run state that survives process death, job queues, resumable checkpoints, webhook-driven resumption, and idempotent tool calls so a retry never double-binds

• Browser agents: Playwright driving real portals end to end, plus diagnosing the one run in ten that fails

• Retrieval architecture: vector retrieval, entity and relationship modeling, chunking and embedding strategy, and keeping facts fresh as they change

• Realtime voice: telephony agents, streaming speech to text and text to speech, sub-second latency, barge-in

• Email agents: thread-aware drafting over full client context without hallucinating a policy detail, leaking another tenant's data, or hitting the wrong recipient

• Application security: authn and authz, secrets handling, prompt-injection defense, threat modeling, and secure tool design

• Postgres with row-level security, and multi-entity schema and query patterns

• AWS (ECS, Lambda, S3), Redis and queues, backpressure, and cost per transaction under load

• TypeScript, Next.js, React

One thing we do not teach on the job. We bind coverage and move real premium, so consequential actions need capability-scoped tools, explicit approval gates, audit trails, reconciliation, and evals that cover them. Prompting technique, orchestration patterns, and context tuning we can teach.

Who does well here

High agency, ships fast, lets the harness write the first draft and then hardens it, owns reliability, debugs intermittent failures, security-minded about tenant isolation, low ego.