TypeScript Engineer — AI/LLM Pipeline & RAG

TypeScript Engineer — AI/LLM Pipeline & RAG (Education SMS Assistant) We run an AI-powered SMS assistant for a fast-growing online school. It's the primary way thousands of parents talk to us — asking about enrollment, programs, their child's progress, and more. You'll be a core engineer on that assistant: improving how it understands messages, retrieves the right information, and replies accurately and on-brand. This is a production system, not a prototype, and you'll work directly with the lead engineer. TWO HARD REQUIREMENTS (please don't apply if you don't meet both): 1. NATIVE ENGLISH FLUENCY. You'll read and reason about real customer conversations every day, judge tone, and write customer-facing copy. Near-native is not enough for this role. 2. CLOSE OVERLAP WITH AUSTIN, TEXAS HOURS (US CENTRAL TIME). This is required, not a preference. You must be able to work a large daily block that overlaps US Central business hours (roughly 9 am–5pm CT). State your timezone and your real overlap in your application. WHAT YOU'LL DO - Improve and debug our retrieval (RAG) pipeline: chunking, embeddings, hybrid (dense + keyword) search, and figuring out WHY the right info did or didn't reach the model - Own our knowledge base system end-to-end: take human-flagged issues, make the fix, reseed embeddings, confirm retrieval, and lock it in with regression tests - Write and maintain evaluation fixtures so quality improvements are provable and don't regress - Work in our multi-agent reply pipeline (routing, prompt composition, structured outputs) - Patch bugs across an async, event-driven backend (queues, webhooks, idempotency) - Read traces and production conversations to diagnose real failures MUST-HAVE SKILLS - Strong TypeScript / Node.js (this is a TypeScript codebase top to bottom — no Python required) - Shipped at least one LLM-powered feature to production (an agent framework, or hand-rolled with the OpenAI / Anthropic / OpenRouter SDKs) - Real RAG experience — and specifically experience DEBUGGING retrieval, not just wiring up "embed docs + top-k" - Comfortable with PostgreSQL / SQL and an ORM (we use Drizzle) - An eval / test-driven mindset — you reach for a regression test, not a manual spot-check - Clear written English and the judgment to evaluate tone and quality in customer-facing replies NICE TO HAVE - Hybrid search / reranking experience (BM25 + dense, RRF) - Experience working closely with a small, fast-moving team - Experience with vector databases, embeddings, etc