AI / ML Engineer

. AI / ML Engineer Contract, roughly 8 months UK based, remote friendly. Reports to the Technical Lead and Solution Architect. Hands-on and production-focused, not research-only. Role overview The AI / ML Engineer builds the agents that make a platform AI-native: document ingestion, onboarding, Analyse and business intelligence, integrity checking and an onboarding agent. You make them accurate, grounded, evaluated and auditable, running safely inside a multi-tenant product, and you ground the interactive avatar in each company's own data. Key responsibilities on this project • Build the custom coded AI agents (not claude co work agents) on top of the structured operational graph, using retrieval-augmented generation with pgvector kept inside PostgreSQL. • Design and maintain the curated, parameterised tool surface and read-only views the agents use, never free-form database access. • Build the evaluation harness that runs each agent against a test set in CI, so behaviour is provable after every change. • Implement the append-only AI audit trail: which agent, and inputs - outputs. • Wire model access through a provider-agnostic layer (Claude Sonnet, GPT-4o and others), with prompt caching, batching and model tiering for cost control. • Keep inference in-region with zero-retention settings and minimal personal data in prompts, for UK GDPR. • Develop a HeyGen avatar and integrate speech to text. • Implement confidence thresholds and human-review queues so low-confidence output is never committed automatically. Essential skills and experience • 5+ years in software engineering, with 2+ years building LLM or ML features in production. You ship, you do not only prototype. • RAG and vector search. Hands-on with retrieval-augmented generation, embeddings, chunking strategies and vector stores (pgvector strongly preferred). • LLM orchestration. LangChain or equivalent, prompt engineering, tool and function calling, and structured output. • Evaluation and safety. You build evaluation sets and guardrails, measure quality objectively, and design against hallucination and prompt injection. • Strong Python. FastAPI, async, and clean, tested service code; comfortable working inside a TypeScript codebase too. • Data fluency. SQL and PostgreSQL, and the judgement to keep retrieval close to the data rather than shipping it to a separate search product. • Provider-agnostic mindset. Experience with more than one LLM provider (Claude, OpenAI) behind an abstraction that avoids lock-in. Desirable • Experience with streaming avatars or speech (HeyGen, Whisper, Deepgram or similar). • AWS Bedrock, in-region inference and zero-retention configurations. • Multi-tenant AI, where each tenant's data and context must stay isolated. • MLOps for AI: prompt and model versioning, and observability of AI behaviour.