AIProjectAssistant

• 

  Solution Diagram

• 

  List of Agents

• 

  AWS Solution Architect outcome

• 

  Invoke SAs lambda

• 

  Invoke PO agent

• 

  PO agent lambda result

• 

  Agent lambda result

• 

  SA Feature Flag

• 

  PO Feature Flag

• 

  NextJS Frontend

• 

  Delivery Director agent invoking Product Owner agent

• 

  Delivery Director agent invoking Solution Architect agent

Inspiration

As consultants we often needed to create technical solution documents and project plans from customer discussions. A lot of time was spent generating these specifications. An AI first solution with agents offering multiple drafts and plans was something that we desired.

Github repo:

Github repo

What it does

The idea is creating a few AI Agents:

• Dee Dee, a Delivery Director
• Poli, a Sr. Product Owner
• Poni, a Jr. Product Owner
• Archie, a Solution Architect very well versed in AWS Cloud services
• Sally, a Solution Architect with deep experience with Open Source services

All agents will rely on Amazon Bedrock Knowledge bases to have access to a transcript of an Event Brain Storming workshop happening between Terazo and EM-Wines, a fictional winery that is looking to implement an online system where they can sell their wines.

Dee Dee, the Delivery Director agent, is going to serve as the "point of contact" for the user. It will be able to answer questions about EM-Wines use case and the implementation of a proposed solution by collaborating with the other agents to create solution designs or a backlog of stories.

To provide a "Human in-the-loop" capability, the solution will rely on LaunchDarkly to implement feature flags and switch between agents whenever the generated content fails user validation.

How we built it

We used Amazon bedrock for creating the agents, lambda functions, S3 buckets, Vector store and invoking the LLM model. Claude Haiku and Sonnet were the LLMs used NextJs was used for a simple user interface. Launch Darly features were used to switch between the agent personas.

Challenges we ran into:

1. We had challenges wording the prompt for accurate output. Small changes to prompt gave varying results
2. Some challenges getting used to Amazon Bedrock.
3. Problems to integrate AWS configuration with Next.JS
4. It seems that agents responses can sometimes time out when invoking them from another agent.

Accomplishments that we're proud of

Was able to learn a lot about implementing agents on Amazon Bedrock. Also this was out first time using LaunchDarkly. We are pleased that we were able to complete the project in time.

What we learned

We learned about the infinite possibilities of AI agents and LLMs. Also how easy it was to integrate with Dark Knightly and enable runtime feature switching

What's next for AIProjectAssistant

We want to test with multiple agents and instead of human in the loop, have AI score the responses from the various agents

Built With

• anthropic
• bedrock
• lambda
• launchdarkly
• next.js
• python
• s3