AI-native requirements engineering on AWS Bedrock
A retrieval-augmented platform that automates requirements analysis, user-story generation and acceptance-criteria drafting.
- Organisation
- JPMorgan Chase
- My role
- Architect and engineering leader
- Period
- Jul 2024 – Nov 2026
The problem
Requirements analysis, user-story writing and acceptance-criteria drafting sit at the front of every delivery cycle on a customer-facing banking platform. Done by hand, they take significant effort and shape the quality of everything that follows, including test design. The goal was to automate this work with generative AI while keeping it grounded in the team’s own source material.
My role
I architected the platform as part of leading AI-native engineering transformation for the platform team, and drove adoption of the horizontal AI initiatives it belonged to across product areas.
Approach and architecture
The platform uses retrieval-augmented generation (RAG). Source material is indexed in OpenSearch Serverless, relevant context is retrieved for each request, and foundation models on AWS Bedrock generate the analysis, user stories and draft acceptance criteria from that context. Grounding generation in retrieved material, rather than relying on the model alone, is what makes the output reviewable against its sources.
Generated artefacts are drafts for people to review. Human review is one of the responsible-AI controls I embedded in the GenAI workflows I led, described below.
Evaluation and governance
Alongside this platform I introduced LLM evaluation and quality-scoring frameworks covering accuracy, groundedness, completeness and hallucination, and embedded responsible-AI controls (human review, data privacy and auditability) in GenAI workflows in partnership with Risk, Compliance and Security. See LLM evaluation and responsible AI controls.
- Sources
- Requirement sourcesTeam source material
- Index
- Vector indexOpenSearch Serverless
- Retrieve
- Relevant contextRetrieved per request
- Generate
- Foundation modelAWS Bedrock
- Review
- Human reviewAnalysis, user stories, acceptance criteria
Documented outcomes
- Automated requirements analysis, user-story generation and acceptance-criteria drafting.
- Formed part of the horizontal AI initiatives adopted by 100+ engineers across product areas.
Team-level results from this role, such as 60% shorter manual regression timelines and about $300K in annual savings, are reported for the role overall and are not attributed to this platform alone.