Rural Reach

• 

  landing page

• 

  about (info) page

• 

  user asks a question from list of suggested topics (prompts)

• 

  chat interface for history

• 

  response from the local GPT

• 

  chat interface for science

• 

  subject (topic) selection for context specific GPT

Inspiration

Having first hand witnessed the difficulties and the marginalization experienced by kids living in rural areas, especially in developing countries such as India, I felt the need to use the paradigm in the distribution of information as an opportunity to provide accessible education to the members of such populations, who often lack access to stable internet connection and therefore miss out on valuable knowledge available on the internet.

What it does

A fully offline fullstack app that uses context specific cloud trained GPT models, available as local deployments through a local fastAPI server.

How we built it

using Electron and React for the frontend, HuggingFace Transformers for the pre-trained GPT models along with Azure Databricks for fine tuning of these models on custom data. The model is made available via a connection to a FastAPI server.

Challenges we ran into

testing different frameworks such as Ray and DeepSpeed for loading and training of the transformer models before settling on azure databricks.

Accomplishments that we're proud of

Developing a functional end to end software that has the potential to provide real value to its users.

What we learned

deploying a fully offline app and the process of downloading pretrained models and training them.

What's next for Rural Reach

More specific fine-tuning of different models with richer datasets, using a more powerful model and better computational resources for faster training. Multi-language support for

Built With

• azure
• databricks
• electron
• fastapi
• huggingface
• react
• three.js