PearlAi

As international students, we faced daily challenges navigating NPU's systems. Information was scattered across multiple platforms, often only in Chinese. The frustration of seeing fellow students struggle with the same issues sparked an idea: what if we could create an AI assistant specifically for NPU?

• Gathered feedback from fellow international students
• Documented common challenges and pain points
• Researched existing solutions and their limitations

Our journey began with testing different language models, each presenting unique challenges:

• First attempt: Mistral 7B - Promising but resource-intensive
• Second try: DeepSeek - Good performance but not optimal
• Final choice: Qwen - Perfect balance of capabilities and efficiency
• Many late nights debugging and optimizing prompts

Working with limited resources, we faced significant technical challenges that required creative solutions:

• Utilized Google Colab for model training
• Dealt with frequent runtime disconnections
• Struggled with long training times and session limits
• Developed workarounds for memory constraints

After months of trials, we finally found our solution:

• Discovered Qwen model - perfect balance of performance and efficiency
• Successfully implemented RAG system
• Overcame context length limitations
• Achieved significant improvements in response quality

The final stretch involved polishing and perfecting:

• Refined prompt engineering for better accuracy
• Improved response time and reliability
• Enhanced context understanding for NPU-specific queries
• Countless hours of testing and adjusting

Currently in testing phase with promising results:

• Positive feedback from initial student testers
• Continuous improvements and refinements
• We welcome feedback to make PearlAi even better