Revolutionizing Education: The Journey of HiSolver’s AI-Powered EdTech Platform

This article was produced as part of the final project for Harvard’s AC215 Fall 2023 course.

Authors: Yan Kaled, Tom Zhang, Tadhg Looram, Mina Lee, Jason Xiang, Nishtha Sardana, Kareema Batool

Project Github Repo — GitHub [Please look into this GitHub for more technical information]

Video — Youtube

Welcome to the fascinating world of HiSolver, where we blend the latest in artificial intelligence with education technology to create an innovative learning experience. In this blog post, we’ll take you through the intricate journey of our project — from its conception to the final product.

The Genesis of HiSolver’s Math Chat

The Team: Spearheaded by a diverse group of talented individuals including Yan Kaled, Tom Zhang, Tadhg Looram, Mina Lee, Jason Xiang, Nishtha Sardana, and Kareema Batool, HiSolver Math Chat represents a melting pot of ideas and expertise.

Our Vision: We set out with a lofty goal — to fine-tune a Large Language Model (LLM) for an animation engine (Python package Manim). Imagine a tool where students can input raw text with mathematical problems and receive step-by-step solutions or hints, not just in text but also as engaging animations!

Crafting the Data Pipeline: A Peek Behind the Curtain

Our project’s backbone is its data pipeline, featuring several key components:

1. Data Collection: Starting with scraping GitHub for repositories related to “manim”, our Python script efficiently gathers Python files storing them in a Google Cloud Storage (GCS) Bucket.
2. Data Pre-processing: We then extract class definition code snippets from these files, prepping them for the next stages.
3. Data Versioning: Using Docker containers, we implement Data Version Control (DVC) on the cloud, ensuring organized and secure data management.
4. Data Labeling: Here’s where human touch comes in. Using Label Studio, users manually provide summaries to the processed code snippets.
5. LLM Fine-tuning: With our labeled data (Python class definitions + manual summaries), we fine-tune a pre-trained LLM (distilgpt2) from HuggingFace, optimized for efficiency and rapid testing via the technique of 16-bit QLoRA.
6. Model Deployment: The fine-tuned LLM is then deployed to Vertex AI, ready to interpret user inputs and generate solutions.
7. Workflow Automation: The entire pipeline is automated with Vertex AI Pipelines, albeit with a pause for manual data labeling.

The Architecture: Designing for Scalability and Efficiency

Our solution architecture is a testament to our commitment to delivering a robust and scalable platform. It encompasses everything from app development, LLM fine-tuning, to user interaction. We utilize a GCS bucket for data storage and Vertex AI for ML workflow automation. The technical architecture further includes a Google Kubernetes Engine (GKE) cluster instance hosting both the backend and frontend.

User Experience: Bringing the Platform to Life

The frontend is a sleek, React-based dashboard with a simple text input field. Users submit their prompts, and the backend, powered by FastAPI, takes over. It renders the Manim package Python code into captivating animations, ensuring a seamless user journey.

Deploying the Dream: HiSolver on the Cloud

Hosted on a Google Kubernetes Engine (GKE) cluster and automated with Ansible, our app’s deployment is as cutting-edge as its design. The Kubernetes cluster configuration is meticulously crafted, balancing computational power with storage needs, and ensuring high availability and security.

A Glimpse into the Future: HiSolver in Action

What we’ve shared so far is just the tip of the iceberg. The final app, deployed at https://hisolver.com, integrates our latest model and extended features, offering an even more enriched user experience. Don’t miss out on exploring it!

Conclusion

HiSolver stands at the intersection of AI and education, redefining how learning is approached. It’s a journey of innovation, teamwork, and relentless pursuit of excellence. We invite you to join us in this revolution, one animated solution at a time.