Senior Ai Systems Engineer— RAG/Agent Architecture, TypeScript (Contract, MVP Build)

We're building a productized AI system for SMB clients that automates internal business workflows using AI agents and a structured data/knowledge layer. We need a senior engineer to build the first working version. Example of the workflow we need built (this is the actual v1 scope): Take a catering company as an example. A client submits an event request. Our system needs to: Send an onboarding form to the client asking for all event requirements (guest count, date, location, dietary needs, menu preferences, budget, etc.) When the response comes back, check it against a defined specification of what information is required If anything is missing or unclear, automatically reply asking for the specific missing details or clarification Follow up (politely, on a schedule) until all required information has been provided Once complete, notify our internal team that the request is fully specified and ready to move forward Write the structured event data into our CRM This requires building: A structured "specification" layer that defines what complete information looks like for a given request type (this is the foundation — think of it as a lightweight knowledge/ontology layer that can later expand to other data types) An AI agent that can parse messy, unstructured client responses, check them against that specification, identify gaps, and draft accurate clarification requests A stateful follow-up loop (not a single email — this needs to track what's outstanding and continue until resolved) A CRM integration to write the finalized data A notification step (Slack or email) to alert our internal team when a request is complete Stack: TypeScript/Node.js, Next.js/React, PostgreSQL (Supabase/pgvector), OpenAI/Anthropic APIs, REST APIs/webhooks/OAuth. Python/FastAPI is a plus. What we need from you: Own this end-to-end: architecture, build, testing, deployment Handle real-world messiness — incomplete responses, ambiguous answers, multiple back-and-forth exchanges Build with reuse in mind — this same system will be adapted for other clients/industries beyond catering, so clean separation between the "specification/rules" and the "workflow engine" matters Reasonable production practices: error handling, retry logic, basic logging — this doesn't need to be over-engineered for v1, but it needs to actually work reliably for a live client This is the first build in a larger system — future phases may include broader knowledge retrieval across documents/calls/emails, more integrations (Xero, Microsoft 365, HubSpot), a client dashboard, and more. We're looking for someone who can grow with this if the initial build goes well, not just complete a one-off task. To apply, include the following — applications without this will not be considered: A specific example of a system you've built that involved an AI agent handling multi-turn, stateful interaction (not a single prompt/response) — e.g., a workflow that had to track incomplete information, follow up, and resolve over multiple exchanges. Link to it if possible (GitHub, live demo, case study), or describe it in enough technical detail that we can tell you actually built it. In 3-5 sentences, describe how you would design the "gap-checking and follow-up" part of the workflow above. Specifically: how would you handle a client response that answers some questions but is ambiguous on others, and how would you prevent the system from sending duplicate or repetitive follow-up messages if the client is slow to respond? What's the trickiest failure mode you'd expect in a system like this, and how would you handle it? (e.g., client gives contradictory information across two follow-ups, client never responds, form data doesn't match any known format, etc.)