The customer’s concierge chatbot currently runs on the Vellum platform and serves hotel guests via a multi-agent architecture (Router, RAG/Knowledge Base, Ticketing, and Sales agents). The primary task is to migrate this system off Vellum while improving its architecture and reliability. This includes migrating the vector store, converting the knowledge base, maintaining existing integrations with the customer’s OPS platform and Property Management System (PMS), and building an evaluation framework to benchmark against the legacy implementation. Future phases with the same customer involve computer vision for property cleanliness inspection and internal productivity assistants for the customer’s staff.
Technology stack: Python, TypeScript, LangChain / LlamaIndex, Vector Store, Node.js, API integrations — OPS platform and Property Management System (PMS), Markdown, Benchmarking and evals.
IN THIS ROLE YOU WILL:
Re-implement the Router logic that dispatches conversations to the correct downstream agent (KDB, Ticketing, or Sales) based on intent classification.
Select and set up an independent vector store to replace the Vellum-hosted RAG document store; migrate existing content.
Maintain the existing integrations with the customer’s OPS platform and Property Management System (PMS) so that the Ticketing and Sales agents continue to operate without disruption.
Convert existing HTML knowledge base articles to Markdown and establish Markdown as the standard going forward; advise on chunking and indexing strategy for optimal LLM retrieval.
Design and run an evaluation framework using a compiled dataset of sample conversations and expected responses; validate that the migrated system meets or exceeds the performance of the legacy Vellum implementation.
Provide a concrete recommendation on hosting architecture: a dedicated Python-based AI server (e.g., using LangChain or LlamaIndex) vs. integrating AI logic directly into the customer’s existing TypeScript OPS backend, with a rationale grounded in scalability and maintainability.
(Future phases) Design a computer vision pipeline that assesses property cleanliness area by area from photos, returns pass/fail ratings, and triggers automated remediation tasks for flagged areas.
WHAT YOU BRING ALONG:
Proven experience building and deploying multi-agent LLM systems in production (not just demos or side projects).
Hands-on work with RAG pipelines: chunking strategies, embedding models, vector stores (e.g., Pinecone, Weaviate, pgvector, Chroma), and retrieval tuning.
Proficiency in Python for AI/ML workloads; familiarity with frameworks such as LangChain,
LlamaIndex, or similar orchestration libraries.
Experience designing and consuming REST APIs; comfort working alongside existing backend codebases.
Ability to design and execute LLM evaluation frameworks (evals, benchmarks, regression testing).
Strong communication skills — you will be making recommendations that non-technical stakeholders need to understand. Ability to clearly communicate discussions output and provide meeting minutes in a timely manner.
Would be a plus:
Experience with TypeScript/Node.js backends (relevant for the integration decision).
Computer vision project experience (relevant for the Inspection AI phase).
Familiarity with hospitality tech (PMS systems, OTA integrations) or similar operational software.
Prior work migrating AI workloads off managed platforms (e.g., Vellum, OpenAI Assistants, Vertex AI Agents).
WHAT WE OFFER:
Professional development:
Highly experienced professional community
Personal development plan and regular Performance Appraisal
Transparent rotation process with an opportunity to switch between different roles, projects, or technology stack
Attendance of professional conferences, meetups, and certifications (coverage upon business needs)
Internal training programs, free SmartTalks, and TechTalks inside the company
Corporate English classes
Well-being:
Flexible sched
