AWS-Mentor-AI
Built an AI assistant using Amazon Bedrock and Nova Pro with tool calling, memory, streaming responses, and knowledge retrieval capabilities.
How I Built AWS Campus Mentor AI Through the AWS Builders Skill Sprint Challenges
When I first started these AWS Builders Skill Sprint Challenges, I thought I would just be building another chatbot. But as I progressed through each challenge, I realized that modern AI systems are much more than just sending prompts to an LLM and getting responses back.
Over the course of five challenges, I built AWS Campus Mentor AI, an AI assistant that evolved from a simple conversational agent into a knowledge-powered assistant capable of using tools, remembering information, streaming responses, and retrieving information from a knowledge base.
This article is less of a tutorial and more of a reflection on what I learned while building the project.
Starting With Challenge 1
The first challenge was straightforward. Build a basic AI agent using Amazon Bedrock, Amazon Nova Pro, and the Strands Agents SDK.
Even though the final application was simple, it was my first time working with an actual agent framework. Before this, I mostly experimented with prompting LLMs directly.
One thing that surprised me was how much difference the system prompt made. Small changes in instructions could completely change how the assistant responded.
At this stage, the assistant could only answer questions and have conversations. It had no memory and couldn't perform actions.
Still, seeing my first Bedrock-powered AI agent run successfully felt satisfying.
Making the Agent Useful With Tools
In Challenge 2, I added tools to the agent.
I implemented:
- Calculator Tool
- Weather Tool
- Age Calculator
This was where the project started becoming interesting.
Initially, I assumed the model would automatically know when to use tools. After testing, I realized that tool descriptions and agent instructions play a huge role in helping the model decide correctly.
There were times where the model tried answering a calculation itself instead of calling the calculator tool.
After a lot of testing and tweaking, the behavior improved significantly.
This challenge helped me understand that building agents is not only about models. It's also about designing the interaction between models and external systems.
Teaching the Agent to Remember
Challenge 3 was probably one of my favorite parts.
The biggest problem with the previous versions was that every conversation started from zero.
If I told the assistant my favorite programming language was Python, it would completely forget it later.
To solve this, I integrated memory capabilities.
The concept sounded simple at first:
"Store information and retrieve it later."
But learning how memory retrieval actually works through embeddings and semantic search was very interesting.
Once memory was working, the assistant felt much more personal.
For example:
User: My favorite programming language is Python.
Later:
User: What's my favorite programming language?
The assistant could retrieve the information and answer correctly.
That moment honestly felt cooler than getting the chatbot working for the first time.
Combining Everything Together
Challenge 4 combined all previous features.
The assistant could now:
- Use tools
- Remember information
- Stream responses
This challenge made me realize how many moving parts exist inside a modern AI system.
A simple user message might trigger:
- Memory retrieval
- Context generation
- Tool selection
- Tool execution
- Response generation
- Memory storage
Seeing all of these pieces work together was rewardling, although debugging them was not always fun.
There were multiple occasions where I spent more time debugging than actually coding.
Building a Knowledge-Powered Assistant
Challenge 5 introduced knowledge retrieval.
Instead of relying only on what the model already knew, the assistant could retrieve information from a custom knowledge source.
This was my introduction to concepts related to Retrieval-Augmented Generation (RAG).
Before this project, I used to think that AI models simply "knew everything."
After working on knowledge retrieval, I understood that many production AI systems actually retrieve information first and then generate answers based on that information.
That approach is much more practical and reliable.
This challenge also made me appreciate how powerful Amazon Bedrock can be when building real-world AI applications.
What I Learned From This Project
Looking back, the biggest lesson was that AI agents are not just LLMs.
A useful AI assistant is usually a combine of these:
- Models
- Memory
- Tools.
- Retrieval Systems
- Context Management.
Every challenge added one of these pieces.
Another thing I learned is that most of the effort goes into connecting components together rather than writing model prompts.
Challenges I Faced
Not everything worked on the first attempt.
Some issues I encountered were;
- Understanding Bedrock configurations
- Debugging tool execution
- Managing memory retrieval
- Making sure retrieved information was relevant
- Handling context correctly
There were several moments where I thought a feature was broken, only to realize I had made a small configuration mistake somewhere.
Those debugging sessions were frustrating at the time, but they taught me a lot.
What's Next?
Although the project is complete for now, I would like to continue improving it.
Some ideas I want to explore are:
- MCP Integration
- AWS Documentation Search
- Advanced RAG Pipelines
- Multi-Agent Architectures
- OpenSearch Integration
I am also interested in exploring how agent systems can be used for education and campus-related applications.
Final Thoughts
The AWS Builders Skill Sprint Challenges provided a structured way for me to learn agent development by actually building projects instead of only reading documentation.
By the end of the journey, I had built an assistant capable of conversations, tool usage, memory management, streaming responses, and knowledge retrieval.
More importantly, I gained a much better understanding of how modern AI systems are designed behind the scenes.
This project definitely changed the way I think about AI agents, and it motivated me to continue learning more about Generative AI and Agentic AI systems.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article