Machine Learning Engineer

Descrial is an AI work-execution platform for alternative investment firms and their portfolio companies. The Machine Learning Engineer will design, build, and evaluate LLM-powered systems, including agentic workflows, retrieval over confidential investment documents, document intelligence, and production evaluation and observability infrastructure. The role also involves developing Python/FastAPI services, optimizing model cost and latency, and operating systems on AWS.

Responsibilities:

• Design and implement multi-step agent workflows in LangGraph / LangChain: state management, checkpoints, tool-calling, and human-in-the-loop patterns
• Build structured-output and prompt systems (Pydantic-validated) for extraction, summarisation, classification, and reasoning over investment documents
• Integrate model providers (Anthropic, OpenAI, Amazon Bedrock) with production discipline, and own model routing and fallback: choosing the right model per task, degrading gracefully when one is unavailable, and absorbing provider model changes without breaking users
• Treat context as a managed resource: what enters a window, compaction across long agent runs, and controlling cost through context design as well as model choice
• Build and consume MCP (Model Context Protocol) tool servers giving agents safe access to internal services, search, and document stores
• Own RAG pipelines end to end: parsing, chunking, embeddings, vector retrieval (Qdrant), reranking, and hybrid search
• Extend knowledge-graph-backed retrieval (GraphRAG) on Neo4j, including Cypher and graph data modelling
• Tune retrieval quality with rigorous relevance evaluation and multi-tenant-safe isolation of indexed content
• Design and run evaluations: golden datasets, LLM-as-judge patterns, and regression gates on prompt or model changes
• Instrument every model call (LangSmith / LangFuse class tooling) for latency, cost, and quality in production — token-cost engineering is a core part of this job
• Apply responsible-AI guardrails: tool-use constraints, output validation, and data-privacy awareness in a UK/GDPR context
• Ship supporting services in Python 3.11+ / FastAPI with PostgreSQL persistence; operate on AWS through CI/CD with quality gates