Academic Projects
- PiMemory: Local LLM Powered Personal Memory Assistant on an Edge Device
- Supervisor: Prof. Bagheri
- Type: Edge AI Systems Project, Retrieval Augmented Generation (RAG), Natural Language Processing (NLP)
- Project Materials: N/A
- This project builds a fully offline RAG assistant that runs on an 8 GB Raspberry Pi 5 with CPU-only LLM inference in C++. Short notes are stored instantly in a SQLite-backed job queue and turned into atomic, dated facts by a background priority scheduler, while questions are answered from BM25-retrieved facts by a locally hosted Qwen3 model. The model fits in memory through 4-bit quantization and KV cache budgeting, and latency is kept low with prefix caching, bounded context windows, retrieval gating, and constrained JSON decoding.
- Benchmarking Prompt Sensitivity in Large Language Models
- Supervisor: Prof. Bagheri
- Type: Research Publication, Natural Language Processing (NLP), Information Retrieval (IR)
- Project Materials: Includes Research Paper, Poster, and GitHub Repository
- This research introduces the Prompt Sensitivity Prediction task and PromptSET dataset, investigating how slight variations in prompt formulation impact LLM performance. Using 11,469 prompts with 9 variations each from TriviaQA and HotpotQA datasets, the study benchmarks multiple approaches including LLM self-evaluation, text classification, and query performance prediction methods. Findings reveal that existing methods struggle with prompt sensitivity prediction, highlighting the need for specialized approaches to help users craft more reliable prompts.
- Query Performance Prediction via LLM-Generated Query Variations
- Supervisor: Prof. Ensan
- Type: Information Retrieval (IR), Natural Language Processing (NLP)
- Project Materials: Includes Project Report, and GitHub Repository
- This project explores Query Performance Prediction (QPP) through Large Language Model–generated query variations. Instead of embedding-space perturbations like ADG-QPP, it uses LLMs (LLaMA 3.1) to produce semantically equivalent queries evaluated with BM25 retrieval. The approach offers interpretable, natural-language robustness estimation and achieves strong correlation performance especially on challenging queries showing LLM based QPP as a viable alternative to dense methods.
- Dataset Generation, and Evaluation for Natural Question Generation
- Supervisor: Prof. An
- Type: Capstone Project, Natural Language Processing (NLP), Natural Question Generation (NQG)
- Project Materials: Includes Project Report, Slides, and GitHub Repository
- This project aimed to create and evaluate a new dataset for natural question generation, emphasizing the practical use of natural-sounding questions in real-world applications. Utilizing state-of-the-art NLP models like BART and T5, the project focused on generating and validating a dataset that effectively aids in natural question generation tasks.
- Social Network Influencer Ranking Based On Link Analysis
- Supervisor: Prof. Papagelis
- Type: Network Analysis
- Project Materials: Includes Project Report and Slides
- This project explores the application of link analysis methods, such as PageRank and weighted PageRank, to rank influencers on social networking services. Focusing on the burgeoning field of digital marketing and online communication, it aims to provide a systematic approach to measuring the influence and reach of key SNS personalities.
Personal Projects
- TTC Real-Time Bus Tracking System
- Supervisor: N/A
- Type: Data Engineering, Real-Time Data Processing
- Project Materials: N/A
- This project develops a real-time application for tracking TTC buses using GPS data from IoT devices. It harnesses Apache NiFi, Kafka, Spark Structured Streaming, and other technologies to process live data, aiming to optimize TTC bus routes.
- Automated Data Pipeline for Transaction Analysis
- Supervisor: N/A
- Type: Data Engineering, Batch Data Processing
- Project Materials: N/A
- This project automates the extraction of transaction data into a cloud-based storage system, followed by its transformation using a scalable compute service. Post-transformation, the data is stored for further analysis and visualized using a business intelligence tool, enabling detailed reporting and analysis of transactional data.
