AI/RAG engineer in document intelligence

Our client is a leading global investment management company headquartered in London. It manages over $228 billion in assets and serves institutional investors, pension funds, wealth managers, and other sophisticated clients worldwide. The firm specializes in quantitative investing, alternative investments, systematic trading strategies, and technology-driven asset management. Data science, machine learning, and AI are core components of its investment and research processes.

• Build the evaluation foundation at the start: convert the client’s existing question bank and diagnostic results into a versioned automated test suite with expert-confirmed outcomes, establish the baseline, and mine historical support records as a second ground-truth set.
• Improve the client’s existing RAG service in place: hybrid retrieval with reranking, structure-aware chunks, context notes, early metadata filters, reliability monitoring, with every change measured against the suite before release.
• Build the metadata and entity extraction pipeline under a governed three-tier schema (universal envelope, versioned per-collection specifications, open discovery tier), with a stratified pilot, confidence-routed human review and coverage dashboards.
• Build the synchronisation plane over the client’s content platform (change notifications, delta polling, scheduled full reviews) shared by retrieval, extraction and downstream views.
• Build document lineage and temporal views: supersession and amendment chains extracted only where stated in the text, a bitemporal effective-terms view, derived document status, and a queryable obligations register, with human verification for high-stakes chains.
• Contribute to the controlled knowledge graph and query orchestration: typed nodes and edges carrying provenance and confidence, routing between structured lookup, filtered retrieval and lineage views, explicit completeness statements, and typed gaps reported as answers.
• Operate the delivered capabilities: releases gated on evaluation results, freshness bounds enforced by withholding stale data, coverage and quality dashboards, corpus health checks reported to document owners.

• 6+ years of production Python development, including 2+ years of LLM and RAG engineering in production: retrieval pipelines, vector stores, structured extraction, and the surrounding operational tooling.
• Strong experience building evaluation harnesses: versioned test suites derived from real question banks, separate scoring of retrieval and answers, scoring where a correct “not found” counts as a pass, and suites wired into delivery as release gates (e.g. Langfuse, RAGAS, DeepEval or similar, plus custom metrics).
• Hybrid retrieval engineering: keyword and semantic search combined, result merging, cross-encoder reranking, tuning against measured baselines, working within a platform-fixed embedding model and index.
• Structure-aware document processing: layout-aware parsing and chunking that keeps tables intact (e.g. Docling, Tika or similar), including OCR handling for scanned documents and multilingual content.
• LLM extraction at scale: schema-driven extraction of attributes, entities, clauses and relationships with per-field confidence, calibrated thresholds and a human review loop (e.g. Label Studio or similar), piloted and measured before scale-out.
• Strong practical experience with AI-assisted tools and agentic workflows, including the ability to identify relevant use cases and integrate AI capabilities into day-to-day engineering practices.
• Strong PostgreSQL: typed relational modelling plus JSONB, schema-as-code with migration tooling, derived views managed as tested transformations.
• Provenance and citation discipline: every extracted fact traceable to its source document and passage; answers that state explicitly when something could not be confirmed.
• Fluent English for written and spoken communication with client teams.
• Будет плюсом: Integration with SharePoint and Microsoft Graph APIs or an equivalent enterprise content platform: change notifications, delta polling, permission metadata; Bitemporal modelling (execution versus effective dates, as-of queries) and document lineage or supersession modelling; Graph engines (e.g. Apache AGE, Neo4j or similar); controlled knowledge graphs with provenance on every element; Legal, contract or fund-documentation domain knowledge, or demonstrated ability to learn a document domain in depth; Permission-aware retrieval: access lists stored in the index, group resolution at query time, denial on uncertainty (the entitlement design is owned by a parallel workstream; this role implements against it); Token-efficiency engineering: corpus preparation, section slicing, cost measurement per query; Exposing capabilities to an assistant platform as MCP tools or skills; day-to-day use of AI coding agents; Experience in financial services or other regulated on-premise environments; client-facing experience

• At Intellias, where technology takes center stage, people always come before processes.
• By creating a comfortable atmosphere in our team, we empower individuals to unlock their true potential and achieve extraordinary results.
• That’s why we offer a range of benefits that support your well-being and charge your professional growth.