[GSoC2026] Reaching out to mentors for "Deep Search AI Assistant on Multimodal Personal Database for AI PC "

[aaartemis5]

Hey @zhaohb @18582088138

I'm Arpit Chaudhary, a third-year Computer Science undergrad from New Delhi, India 👋

A bit about what I have been upto in last year :

I was awarded National Rank 1 in India for my ML research by the Prime Minister of India and the Chairman of ISRO (Indian Space Research Organisation) on National Space Day for research on a novel physics-aware neural network for solar event prediction (CME) using plasma data, bypassing traditional magnetic field methods entirely.
I was also one of the youngest undergraduates to present research poster at the joint ESA(European Space Agency) -ISRO International Heliophysics Workshop.

On the industry side, I previously worked at an IoT-based startup where I built a production-level RAG-based Root Cause Analysis tool for industrial logs, a Gen-AI assisted OCR pipeline*that reduced manual logbook entry time by 75%, and anomaly detection models for long-term system monitoring.

While building RAG based RCA (Root cause analysis tool ) I also worked with a lot of poorly scanned images from really old books which cant be turned into database chunks easily, a multimodal personal deep search AI , should be able to go through the content of those pages in images that we click or that are scanned and sstore it in database as well to query further , to get the best out of its abilities , then it will be embedding all of that information in those pages as well and those can be queried as well further.( including tables , diagrams andd much more)

I'm very interested in the "Deep Search AI Assistant on Multimodal Personal Database for AI PC " , this is a personally motivational topic for me to build in because I have always wanted a personal Local AI assistant.

-Beyond my professional internships, I have been developing a personal passion project called LogVision (demo attached). It is a privacy-first, multimodal RAG system designed to bridge the gap between continuous video feeds and natural language querying.

The core of the system is a flexible Privacy-Tiered Indexing architecture that I believe is directly applicable to the Deep Search AI Assistant:

1. Privacy-First Mode: In this state, the system uses YOLO-based Computer Vision to analyze video in real-time and extract only text-based event logs (e.g., 'Person entered the room at 2:00 PM'). These text embeddings are stored in a local RAG database, allowing users to query their history without ever saving a single frame of raw video.
   

2. Multimodal Mode: If the user opts for full functionality, the system indexes both the text descriptions and the corresponding video clips. When a user asks a question, the Multimodal RAG retrieves the most relevant text answers alongside the top matching video evidence.
   

This is the attached demo of what i have been working on (this is just a demo , i am working on improving a lot of things )

WhatsApp.Video.2026-02-28.at.23.53.24.mp4

Without the dual rag functionality , it only goes for the text rag made from videos , without storing videos , maintaining privacy.
Using CLIP , yolo , Re-id , qwen( can make this a local model too in future)

I am very interested to work on Deep Search AI assistant on Multimodal personal database and I think all of my experience will help me deliver something that will be truly effective.

Looking forward to discuss more and develop more!

Arpit Chaudhary