AWS Elevate

Agentic AI Architecture
React JS frontend via AWS Amplify, API Gateway, Lambda, and Planner for quiz, QnA, recommender, and knowledge check.

Inspiration

When someone starts on their AWS learning or certification journey they usually get plagued with analysis-paralysis. Which cert to pick? How much effort will it take? How to start preparing? That's exactly why we built this Agentic AI solution to help learners on this journey.

What it does

- It understands the user's profile and recommends a certificate to pursue
- It uses RAG to provide contextual grounded answers related to the certification (e.g. cost, duration, etc)
- It provides a personalized Quiz module that creates mock questions on the fly and provides customized feedback
- It has an in-built Tutor as well to provide goal-oriented teaching.

How we built it

- We leveraged Amazon Bedrock Agents for the Agentic layer
- We used DynamoDB for implementing the interim trace-events that Bedrock Agents emits (reasoning, tool calls, agent collaboration, etc)
- Finally we used a React.js application to provide a good UX to interact with the Agents that uses polling mechanism to provide informational messages to the user.

Challenges we ran into

- Smaller models were not behaving appropriately. We solved it by switching to the higher capacity models e.g. Nova Premier
- In-depth reasoning was leading to higher latency. We worked around this by showing "behind-the-scenes" agent messages to the user.

Accomplishments that we're proud of

- Built a useful system that we ourselves are going to use.
- Built a UI that is really cool to interact with and gives a peek under the hood of how agents are actually responding to the user's requests.

What we learned

- The multi-step events emitted by Bedrock Agents does allow for a lot of flexibility. 
- The yaml based binding of tools to agents is a neat way of attaching tools.

What's next for AWS Elevate

- We plan to add memory related to topics learnt
- Connect with email or IM to follow the user on a regular schedule (e.g. 2-min learning capsule in the user's inbox every day)