LLM evaluation and responsible AI controls
Quality-scoring frameworks for accuracy, groundedness, completeness and hallucination, with human review, privacy and auditability built in.
- Organisation
- JPMorgan Chase
- My role
- Engineering leader
- Period
- Jul 2024 – Nov 2026
The problem
Generative AI output can be fluent and wrong. In a regulated banking environment, teams adopting GenAI need a consistent way to judge whether output is good enough to use, and controls that satisfy Risk, Compliance and Security, without every team inventing its own.
My role
I introduced the evaluation and quality-scoring frameworks, set the reusable standards, and partnered with Risk, Compliance and Security on the responsible-AI controls. I also spoke at JPMorgan Chase engineering forums on AI-enabled tooling.
Approach and architecture
Evaluation. Output is scored on four dimensions:
- Accuracy: is it correct?
- Groundedness: is it supported by the source material it was given?
- Completeness: does it cover what was asked?
- Hallucination: does it introduce content that is not supported at all?
Making these dimensions explicit and shared turned them into reusable AI quality standards across teams, rather than one-off judgements.
Controls. Three responsible-AI controls were embedded in GenAI workflows:
- Human review of generated output before it is used.
- Data privacy for the information the workflows handle.
- Auditability, so decisions and outputs can be traced afterwards.
Evaluation and governance
The controls were designed in partnership with Risk, Compliance and Security, so governance was part of how the workflows were built rather than a review at the end. This draws on earlier work in regulated medical-device software, where I worked to IEC 62304 and ISO 13485.
- Output
- Generated outputFrom a GenAI workflow
- Evaluate
- Accuracy
- Groundedness
- Completeness
- Hallucination
- Control
- Human review
- Data privacy
- Auditability
- Standard
- Reusable AI quality standardsShared across teams
Documented outcomes
- Set reusable AI quality standards across teams.
- Embedded human review, data privacy and auditability controls in GenAI workflows.
- Shared the approach at JPMorgan Chase engineering forums on AI-enabled tooling.