Selected AI & ML Projects
Academic Accommodation Recommendation Engine
Technologies: Python, PyTorch, LangChain, Azure AI, Prompt Engineering, LM Studio
- Designed and built a Retrieval-Augmented Generation (RAG) system to recommend personalized academic accommodations for students.
- Integrated Azure AI Services for secure, scalable deployment.
- Improved recommendation relevance by 12% compared to rule-based methods; provides interpretability by explaining rationale.
- Under deployment by Carleton University (Fall 2025).
Banking Customer Service Chatbot with Agentic RAG
Technologies: RAG, LlamaIndex, Open-Source LLMs, LangGraph, LangChain, Vector Databases
- Built and deployed an RAG-based chatbot to handle general banking queries, reducing call center workload by 23%.
- Implemented context-aware retrieval to ensure accurate and compliant answers.
- Deployed by a Private Bank, Iran.
Deep Learning-Based ICD Coding of Clinical Records
Technologies: Python, TensorFlow, CNN, RNN, Hierarchical Classification
- Evaluated and benchmarked deep learning models for automated multi-label classification of discharge summaries.
- Integrated hierarchical ICD code structure into model design, boosting accuracy.
- Applied few-shot learning with category description vectors.
- Performance improved by 7% over traditional ML/DL methods.
- Deployed as part of HMS at University Hospital, Leuven, Belgium.
Automated Skill Network Generator for Job Descriptions
Technologies: AWS SageMaker, Python, TensorFlow, LIME, spaCy, Multi-Task Learning
- Built an end-to-end AWS SageMaker pipeline to parse job descriptions and build skill graphs.
- Applied ensemble learning and transformer-based NER for improved skill extraction.
- Annotated custom dataset of 22,000 job postings across 22 IT roles.
- Incorporated LIME explainability for interpretable predictions.
- Achieved F1-scores of 72% (technical) and 67% (non-technical).
- Deployed as part of the Skills Cloud system by Begreator startup, Belgium.
Prompt-Optimized Text Summarization Evaluator
Technologies: HuggingFace Transformers, Prompt Engineering, LangChain, LlamaIndex
- Designed a summary evaluation framework using Chain-of-Thought and Tree-of-Thoughts prompting.
- Achieved more consistent evaluation than ROUGE/BLEU with semantic similarity checks.
- Showed feasibility of evaluating summaries without ground truth.
- Won the Best Solution Award at the 4th ACM Workshop on NLP Systems Evaluation.
Multimodal Visual Question Answering System
Technologies: BLIP-2, PEFT, LORA, Vision-Language Models, Multimodal AI, LangChain
- Built a parameter-efficient fine-tuned VQA pipeline combining vision-language models with LLM reasoning.
- Developed a dual adapter architecture (CLIP-Adapter + Task-Residual).
- Integrated LLM to generate attribute-specific prompts and context-aware descriptions.
- Achieved SOTA results on 11 benchmark datasets and 5 few-shot settings.
- Applications include medical imaging and personalized AI.
- Published in International Journal of Computational Intelligence Systems.