AI Expert for Contact Center Assistant

# AI Expert for Contact Center Assistant Short version for the Upwork listing. The full technical brief (see `docs/brief-expert-ai.md`) is provided as an attachment. --- ## Suggested job title **AI Engineer — Agentic RAG + Voice for Production Contact Center (Roomonitor)** Alternatives: - Senior AI Engineer — Agentic Chatbot + Tool Use + Voice - LLM Expert — Production RAG → Agentic Migration + HubSpot & Voice Integration ## Category AI & Machine Learning → Chatbot Development / LLM Development ## Experience level Expert ## Project length 3–6 months (phased) ## Hourly / Fixed Fixed price per phase (recommended). Open to hourly for initial discovery. --- ## Job description (paste this) **⚠️ Before you apply — please read** We need results **in days, not weeks**. This is not a learning opportunity — we're looking for a seasoned expert who has already shipped this exact type of work multiple times and can hit the ground running. - **Do NOT apply** if you don't have direct, hands-on production experience with agentic LLM architectures (Anthropic tool_use, OpenAI function calling, or equivalent frameworks). - **Do NOT apply** if you can't start this week and dedicate significant hours to move fast. - **Do NOT apply** if your portfolio is only demos, tutorials, or courses — we need proof of production systems in the wild. Proposals without portfolio evidence of shipped agentic/RAG systems will be dismissed. Proposals that don't address the timeline explicitly will be dismissed. --- **About us** Roomonitor is a Spanish company managing 20,000+ vacation-rental apartments across Spain. We've built **Agente-RM**, an AI contact center: an LLM-powered assistant that resolves guest incidents by following documented protocols and using real apartment data (access codes, WiFi, appliances, alarm systems). It's live in production at agente.roomonitor.com and supports two audiences: guests (with sensitive info hidden) and human agents (with full visibility). Current stack: Node/TypeScript + PostgreSQL 16 + **pgvector**, Anthropic **Claude Sonnet**, OpenAI embeddings, React 19, Docker on Ubuntu. Real production data: 1,288 companies, 10,575 buildings, 20,861 apartments, 3,919 protocols with embeddings. **What we need** We're looking for an experienced AI engineer to take the assistant from **one-shot RAG** to a **production-grade agentic system** with three phased goals: 1. **Accuracy (core)** — improved retrieval (re-ranking, hybrid search, hnsw index, query reformulation), **migration to agentic architecture (tool use)** so the assistant fetches data on demand, a proper **evaluation harness** with test set + metrics, and robust handling of multi-language and dirty data. 2. **HubSpot integration** — connect the AI to HubSpot's chat as the customer support channel. 3. **Voice / telephony** — the AI answers phone calls (STT → LLM → TTS) with conversational latency and human handoff. **Additionally (transversal)**: implement **structural security barriers per role**. Today guest-vs-agent separation is prompt-only and vulnerable to prompt injection. We want enforcement at the data/tool layer (defense in depth). **What we provide** - Private TypeScript repo, documented, with a reproducible local environment (Docker + Postgres + production data replica) - Production server access - OpenAI + Anthropic API keys already in use - **Full technical brief attached** with architecture, data model, current limitations, and scope per phase **What we need in your proposal** - Architecture proposal for the phased plan - Effort estimate per phase **in days**, plus availability starting this week - Success metrics you'd propose for "accurate responses" and how you'd measure them - Recommended stack additions (vector index, reranker, voice platform) with cost impact - Collaboration model (turnkey vs. iterative) and what you'd need from us - **Links to 2–3 production agentic/RAG systems you've shipped**, with impact numbers if possible --- ## Required skills / tags - LLM / Anthropic Claude / OpenAI - Retrieval-Augmented Generation (RAG) - Agentic AI / Tool Use / Function Calling - pgvector / Vector databases / Semantic search - TypeScript / Node.js - PostgreSQL - LLM evaluation (test sets, metrics, hallucination detection) - Prompt engineering - Voice AI / STT / TTS (nice-to-have for now, mandatory later) - HubSpot integration (nice-to-have) - LLM security / Prompt injection defense --- ## Preferred qualifications - **3+ years shipping production LLM systems** (not prototypes, not demos, not courses) - **Shipped agentic systems with tool use** in production (Anthropic tool_use, OpenAI function calling, or equivalent) — this is a hard requirement, not "nice to have" - **Available immediately** — starting this week, with significant hours committed - Portfolio of RAG-to-agent migrations with measurable accuracy gains - Familiarity with evaluation frameworks (LangSmith, Braintrust, LlamaIndex evals, custom) - Comfortable in a real production codebase (TypeScript strongly preferred)