Hassaan Muhammad Khan,
an AI & Machine Learning
Researcher from Pakistan!

This project was done in collaboration with MBZUAI, as a research intern under professor Zhiqiang Shen. This research empowers the small language models to perform contextually well through knowledge distillation from a subsequently larger language model. Check out the detailed evaluation on different benchmarks getting upto 27% increase in accuracy. Project Link.

As the name suggests, this state of the art robot uses Lidar technology to map the area and navigate autonomously within it. Fun-fact: This robot has a fine-tuned LLM deployed on server that helps it to interact along with RAG pipeline. Check out more on the github repo: Project Link.

This research was done during my time at Additech-Sim and is one of the various industrial research projects completed. The research focuses on improving existing pressure values within a certain boundary conditions through a custom stable-baseline3 RL environment. Pretrained and optimized supervised learning model was used for inference. Check out more on the link: Project Link.

This project was done as a part of DOAZ, to analyze Basic RAG, Corrective RAG and Graph RAG, using end to end langchain and Neo4j implementation. 5 Pdfs were used to create FAISS vector DB using large embedding model from openai which was used for RAG and CRAG. GraphRAG used knowledge graph saved on server and cypherquery through neo4j for efficient retrieval. Detailed comparision saved on excel sheets. Project Link.

The project makes use of pretrained VIT Transformer fine-tuned on FER and AffectNet datasets for facial emotion recognition. CNN was also trained with different variations of epochs on same datasets and analysis was done based on validation accuracy as highlighted in the repository, do check it out! Project Link.

Our project, conducted under Dr. Zhiqiang Shen during UGRIP at MBZUAI, was titled as "Optimizing Prompts for Foundation Models" to reduce hallucination. We curated a benchmark dataset of 25k questions across ~60 topics like engineering, philosophy, and history. Additionally, we developed a web application to collect human preferences and assess the correctness of responses before and after applying 26 guiding principles from VILA LAB. This preference data is crucial for future preference-based optimization techniques, enhancing the accuracy and reliability of AI-generated responses Project Link.