compliance regulatory agent

multi agent systems to help streamline the work for compliance team in financial industry - help with compliance review, ediscovery, archiving and etc.

Comment

Inspiration

Everyday compliance officers in a financial services company get drowned by the amount of compliance pipeline workflow and reviews on communications data between their employees and clients to ensure that all the communications stay compliant with all the regulatory rules such as SEC, FINRA, FCA and etc. They need to search through hundreds thousands of emails not really knowing what to look for, they need to find the needle in a haystack when there is a communication message that might not be compliant and communications data are only going to grow bigger and bigger.

What it does

This multi-agent system helps compliance officers do their "Step 1" task for their daily job on discovering their archived communication messages and also the agents help conduct preliminary reviews and flag the more important ones for human to review it. This help streamline their everyday workflow so that they can focus on what's important and not do the tiny bits not knowing where to start.

How we built it

For the agents framework we definitely used Google ADK which has a root orchestrator agent that will delegate tasks to it's sub agents: case_officer_agent, supervisor_agent, compliance_analyst_agent.

  • case_officer_agent & supervisor_agent will use archivist_agent as an AgentTool to retrieve the archived communication messages from Google Cloud Storage requested by the user query.
  • After the agents received the returned data, they will then use the MCP Toolbox Databases to store the data in AlloyDB for operational purposes.
  • compliance_analyst_agent will be used when the users want the agent to do a preliminary review on all the messages that is in the queue.

Challenges we ran into

  • Searching through the archived communication data with Vertex AI similarity search at first but found out it won't be able to do metadata filtering such as dates or sender ## Accomplishments that we're proud of
  • Managed to create an end-to-end proof-of-concept project with Google ADK after just learning about it for 2 weeks.
  • Implemented Vertex AI Search to do metadata filtering.

What we learned

  • Implementing end-to-end agentic ai applications with Google ADK + FastAPI + AlloyDB + MCP Toolbox Databases
  • How to build multi-agent system workflow
  • How MCP works.

What's next for compliance regulatory agent

  • Another agent to parse through uploaded documents about a legal case and search through communication messages that might be relevant for legal purposes.
  • Add another source of communication messages, maybe even multimodal such as calls, video meetings, MSTeams/Slack chat with images.
  • Improve knowledge search with Graph-RAG so that the system is able to know which thread is this message referring to.
  • Add agent's reasoning for each messages reviewed
  • Present agent's CoT process

Built With

Share this project:

Updates