Oxford Risk: Unlocking GenAI Foundations with ADAPT

The snapshot

About Oxford Risk

Oxford Risk is a UK-based behavioural finance fintech serving wealth managers, banks, pension providers, and robo-advisers.

The company provides solutions that support investment and pension decision-making, grounded in rigorous academic research and a deep understanding of client psychology and financial behaviour.

With strong growth and partnerships with leading global wealth managers, Oxford Risk maintains an award-winning team of behavioural scientists, quantitative finance experts, and technologists.

The Challenge

Oxford Risk faced a critical challenge in unlocking value from its extensive repository of behavioural and financial content:

• Large unstructured knowledge base – Research papers, behavioural studies, and proprietary content accumulated over the years.
  
• Indexing uncertainty – Identifying the optimal methods to structure documents for LLM access.
  
• Diverse content types – Behavioural assets required different chunking and embedding strategies.
  
• Precision vs metrics gap – Traditional retrieval scores did not fully reflect real-world GenAI relevance.
  
• Need for rapid validation – Required empirical insights quickly to avoid long planning cycles and maintain competitive advantage.

The Solution

Devoteam partnered with Oxford Risk to deliver a structured two-week strategic validation using AWS native services and the ADAPT methodology.

Rapid Prototype Development

Workshops focused on narrowing the scope to data preparation for GenAI, identifying the need to experiment with RAG indexing and vector store configurations across behavioural content types.

GenAI Architecture Design

Devoteam designed a modular, scalable architecture using AWS services:

• AWS Bedrock for semantic analysis and embeddings
• S3 Vector Store for retrieval
• DynamoDB for metadata and state management
• AWS Lambda for contextual document transformation
  

Three separate knowledge bases were created, each testing a different chunking strategy:

• Default fixed-length chunking
• Semantic chunking

Contextual enhancement with document-level enrichment

Accelerated Validation

Multiple queries were executed across all knowledge bases and evaluated using retrieval scoring and LLM-based relevance assessment.

The results highlighted a key insight: while fixed-length chunking performed best on retrieval metrics, contextual and semantic approaches delivered superior real-world relevance.

Embedding visualisation further validated metadata enrichment, showing clear segmentation of documents by domain and confirming effective routing of behavioural, research, and proprietary content through the pipeline.

The Value Delivered

Rapid empirical validation
Oxford Risk gained actionable insights within weeks, significantly reducing the time required for architectural decision-making.

Improved AI relevance
Contextual and semantic approaches enhanced the quality and usefulness of GenAI outputs.

Scalable foundation
A modular architecture was established, supporting future enhancements such as hybrid retrieval, re-rankers, and graph-based RAG.

Governed behavioural AI vision
The solution contributed to a broader strategy for a unified knowledge base and behavioural AI layer across the platform.

AWS funding support
Access to AWS’s Last Mile Accelerator Program provided additional funding to support the transition to full-scale production.

Why Devoteam?