Chat with my Resume - Building a Doppelganger Part 1

🤖 Just built a personal AI assistant that's smarter than me!🤓

Well that's a bit far fetched, but yes, a side project I loved working on. Curious? Try it out and let's explore the possibilities together! Link here🌐:https://lnkd.in/gqNnA5AN

Since I heard the revolutionary 'ChatGPT' it has been my wish to work on fine tuning an AI to replicate my own style and way of answering, maybe respond to emails, texts, or queries even in my absence (or attend meetings for me 👀). However this project is far away from that, but it sure is a beginning and a fun project.

Leveraging the RAG system, Llama Index, and a LLM at backend, this bot offers answers by indexing my resume and key documents. It uses Pinecone as the vector database option. Special commands like "/download" provide the complete resume for people to download.

Key Learnings and Next few steps🗝️:
• This bot is no where near to perfect, especially incase of inference time. Even though the LLM inference is on par, the time taken to index the queries can be improved. Backend optimizations like parallel processing of components, or using a distilled model fine tuning to mimic user's personality are essential.

• Fine-tuning the model to handle the specific language and lingo we use, such as "Hinglish," requires a robust database of my chats and interactions.

• And majorly, the GPU availability and cost plays a vital role in such projects. Other issues like scalable infra implementation, model quantization for speed, batch processing, etc are to dos for the future.

Yet, it is a small exciting victory and I'm eager to continue refining this project to better reflect my persona. Stay tuned for more updates!