TypeScript / LLM Engineer — Conversational AI Agent (Sports Betting)

We’re building a production conversational AI that texts members over iMessage/SMS about sports betting — picks, odds, live scores, and ongoing banter. The agent has a distinct voice, tool calling into our own data, multi-turn memory, and a golden-eval loop. We need an engineer to deepen conversation quality, tool grounding, and the reply pipeline. This is backend / LLM product work, not a greenfield chatbot. The system already ships; you’ll extend and harden it. What you’ll work on Conversation quality — brevity, honesty, temporal grounding (“today” vs stale memory), multi-turn coherence, group-chat behavior Tool use & grounding — odds board, live scores, player lookup, recall tools; stop invented numbers; keep web search for gaps only Persona / prompts — system voice, prompt versioning, latency-aware reply lanes (fast vs full) Memory & personalization — thread memory, communication style, member model; privacy-safe updates Eval & regression — golden conversation scenarios + LLM judge; unit tests for pure helpers Inbound pipeline — webhooks → gates/routing → context assembly → agent loop → outbound bubbles; acks, rate limits, failure fallbacks Stack TypeScript on Deno (Supabase Edge Functions) Linq for messaging Postgres / Supabase Anthropic Claude (Sonnet for persona + tools; Haiku for routing/judges) iMessage/SMS webhook integrations Deno unit tests + scenario-based persona evals Primary surface is the conversation backend. iOS companion app exists; not the main focus. Requirements Strong TypeScript (async pipelines, testable pure helpers) Shipped LLM agents: system prompts, tool schemas, agent loops, eval harnesses Comfortable with Supabase/Postgres and serverless/edge functions Experience with multi-turn chat: routing, memory, latency UX Sports/betting domain familiarity (odds, slates, scores) is a strong plus Solid testing habits (unit + scenario evals) Nice to have SMS/iMessage or messaging webhook systems Prompt caching, voice/style systems, LLM-as-judge evals Deno experience Engagement Contract / part-time or full-time — open to the right fit Async-friendly; clear PRs against an existing codebase You’ll get access to the conversation modules, eval harness, and recent voice/tool work To apply Link to a shipped conversational or tool-calling LLM project (repo or product) Brief note on how you’ve measured or improved reply quality (evals, human review, etc.) Hours/week available and typical turnaround for a focused PR Your rate Please also type "penguin" in your letter.