Product, architecture, engineering, DevOps, QA and localization
01 Context
The work, at a glance.
As Lead AI Engineer at an enterprise compliance platform, I lead the agentic AI workflows that keep its compliance forms current. When a regulation changes, the platform finds the affected content, drafts updates and sends them to subject-matter experts as change requests. It runs in production.
PythonAWSAmazon BedrockLangSmithMongoDBNext.js
02 The challenge
What needed to change.
Regulations change constantly, and a single change can affect content across many forms. Finding it by hand is slow and error-prone, and compliance content can’t be rewritten by a model without an expert signing off.
Find every piece of content a regulatory change affects
Draft accurate updates without touching what didn’t need to change
Keep experts in control of every change
Keep translated forms in step with their source
03 The approach
Decisions that moved it forward.
The platform is built around one rule: AI proposes, experts decide. I led the AI side, from linking content to regulations through to drafting and checking updates.
01
Change Propagation
Turns a regulatory change into proposed updates for the content it affects.
Finds the questions and forms linked to the changed regulation
Drafts updates with large language models on Amazon Bedrock
Checks every draft before it reaches a person, and drops changes that add nothing
Sends what passes to experts as change requests to approve, edit or reject
Shows reviewers the published and proposed versions side by side, with the changes highlighted and the source rule in view
Labels every suggestion as AI-generated and lets reviewers rate it
02
Linking Content to Regulations
Connects each question to the regulations behind it, so changes can be traced in both directions.
AI-assisted linking that experts can inspect and correct
A searchable, versioned regulatory knowledge base
03
Translation and Quality
Applies the same review rules to translated forms and to the AI work itself.
Translations go to experts for approval before they are applied
Tracing and evaluations to monitor and improve AI output
04 The outcome
What the work made possible.
The platform is in production. Regulatory changes now reach experts as reviewable updates instead of a manual search through every form.
One regulatory change, end to end
A regulation changes: the inspection interval drops from 12 months to 6.
A regulation changes: the inspection interval drops from 12 months to 6.
The platform finds every form with a question built on that rule.
For each affected form, AI drafts an edit to the linked question.
Checks make sure the edit comes from the change and keeps the form valid.
Edits that only reword things are dropped.
What survives becomes a change request.
An expert approves, edits or rejects it. Nothing applies itself.
Illustrative: the rule, question and counts are made up.
Status
In production
Role
Lead AI Engineer
Review
Every AI change approved by an expert
Scope
Change propagation, linking and translation
Experts decide every change. The AI finds, drafts and checks.
Each update traces back to the regulation that caused it.
Translation follows the same propose-and-approve flow.
