RiskFocus — Risk Summariser
Multi-document risk extraction that surfaces the connections between documents — where they reinforce, qualify, or contradict each other — with traceable lineage back to every source.
- Domain
- Risk assessment
- Lineage
- Multi-document
The problem
Risk rarely sits inside a single document. It emerges from how documents relate — an exposure named in one file, qualified in a second, contradicted by a third. Extracting each document in isolation misses exactly the thing an assessor needs, and a summary nobody can trace back to a source is a summary nobody can defend.
The approach
RiskFocus shares its foundations with the document summarizer but targets a different problem — risk extraction across a document set, and the relationships within it:
- Multi-document risk extraction, identifying risk-relevant content across a whole set rather than file by file.
- Connections between documents, surfacing where they reinforce, qualify, or contradict one another — the part that isn’t visible from any single file.
- Data lineage, so every element traces back to the document it came from. Without this, an assessor cannot defend a conclusion.
- Risk factor management by line of business, letting each business line define the factors relevant to it instead of sharing one flat taxonomy.
- Knowledge-base-backed prompts, grounding generation in curated domain knowledge rather than leaving the model to improvise risk language.
- Q&A workflows for guided decision-making, so assessors could interrogate the evidence instead of only reading a static summary.
Delivered on AWS with Bedrock for inference, Step Functions for orchestration, and Lambda plus ECS for the processing tiers.
The outcome
Assessors moved from reading generated summaries to working with them — following any statement back to its source, seeing how the documents relate to each other, tuning the risk lens to their own line of business, and asking follow-up questions against the underlying evidence.