I'm a Computer Science undergraduate at Concordia University, specializing in Data Science and interested in AI/ML engineering, LLM systems, agentic workflows, retrieval-augmented generation, and backend software development. Through two internships and hands-on projects, I've worked on data preparation, model training, hybrid retrieval, AI orchestration, performance evaluation, inference optimization, and deployment.

I enjoy building end-to-end systems that people can actually use, with an emphasis on reliability, measurable performance, and production-minded engineering.

• Built a Python/FastAPI multi-agent CRM assistant for support, knowledge retrieval, and automation, achieving ~4s P95 latency with 30+ concurrent sessions.
• Engineered user-isolated hybrid RAG with BM25, ChromaDB, BGE, reciprocal rank fusion (RRF), and cross-encoder reranking, incorporating reliability safeguards and Docker-based LLM serving.
• Designed and developed 20+ reusable Langflow nodes for modular AI orchestration, workflow integration, and automated task execution.
• Conducted A/B experiments across 200+ clients, analyzing system performance and comparative outcomes to inform product improvements.

• Built a Top-K movie recommendation system using 100K+ ratings across 9K+ movies, benchmarking Factorization Machines vs. SVD++ with 10-fold cross-validation.
• Optimized ONNX Runtime batch inference to approximately 0.7s P95 latency under 50+ concurrent sessions.
• Cleaned and validated 50K+ NLP samples, achieving 98%+ data-quality pass rates; implemented automated testing and P50/P95 monitoring, reducing P95 latency by ~18%.

Built a medical-information Graph-RAG prototype using BM25, FAISS, Neo4j/Cypher, cross-encoder reranking, HyDE, query expansion, and context compression, with structured output validation.

Results: 0.92 faithfulness · 0.81 answer relevancy · 0.89 context precision · 0.74 context recall (RAGAS, 200 queries) · ~1.1s P95 TTFT
Stack: Python FAISS PostgreSQL Neo4j Qwen 2.5 Ollama Docker

Developed an end-to-end reinforcement learning from human feedback (RLHF) pipeline with supervised fine-tuning, a RoBERTa reward model, and custom actor-critic PPO, using UltraFeedback, margin-ranking loss, and KL regularization. Deployed interactive inference with Streamlit and Docker.

Results: 546 PPO wins vs. 53 SFT wins · ~91% PPO win rate in LLM-as-a-judge comparisons
Stack: Python PyTorch Hugging Face GPT-2 RoBERTa PPO React Native Node.js Streamlit Docker

Fine-tuned Sentence-BERT with LoRA on 404K Quora question pairs, using contrastive, triplet, and classification objectives. Combined semantic embeddings with 13 lexical features in a 1,000-tree LightGBM GOSS ensemble.

Results: 0.82 ROC-AUC
Stack: Python Sentence-BERT LoRA LightGBM React Native TypeScript AWS DynamoDB

Built a five-class sentiment analysis pipeline using PySpark, FastText, and BiGRU/BiLSTM networks, with containerized inference. Compared model performance using 10-fold cross-validation.

Results: BiGRU: 83.1% ± 12.5% accuracy · BiLSTM: 82.4% ± 7.8% accuracy
Stack: Python TensorFlow/Keras PySpark FastText BiGRU BiLSTM CUDA Docker

• 🎮 Deep Reinforcement Learning for 10×10 GridWorld — Benchmarked DQN, Double DQN, and Deep Expected SARSA for obstacle-aware navigation across 3 evaluation maps, with 22,000+ state transitions and a 100-state / 4-action environment.
• 🌙 Sleep Disorder Multi-Class Prediction — Compared Logistic Regression, polynomial-kernel SVM, and Gradient Boosting, with preprocessing, feature engineering, PCA, and cross-validation; achieved 0.88 macro-F1 with Gradient Boosting.

• Languages: Python, C/C++, Java, JavaScript, TypeScript, SQL, HTML5/CSS
• AI & ML: PyTorch, TensorFlow/Keras, CrewAI, LangChain, LlamaIndex, RAG, tool calling, MCP, vLLM, CUDA Graphs
• Backend & Frameworks: REST APIs, FastAPI, PostgreSQL, MongoDB, Redis, Kafka
• Cloud & DevOps: AWS (EC2, Lambda, RDS, S3), Azure, GCP, Docker, Kubernetes, Linux, Git, GitHub, CI/CD
• Hands-on ML Systems: Hybrid retrieval, cross-encoder reranking, ONNX Runtime inference, A/B testing, load testing, P50/P95 monitoring

• Peer Tutor Volunteer — Google Developer Groups on Campus, FPT University (Feb 2026 – May 2026)
  Helped students strengthen programming, mathematics, and problem-solving skills through peer discussions, coursework review, and small-group study sessions.
• Event Support Volunteer — Concordia University Economics Student Society (Sep 2026)
  Assisted with event setup, attendee check-in, participant guidance, and logistics.
• Campus Open Day Volunteer — FPT University (Sep 2024 – Sep 2025)
  Supported visitor registration, campus tours, and coordination for prospective students and families.
• Community Volunteer — Local Student Volunteer Group (Nov 2024 – Dec 2024)
  Participated in community cleanup, litter collection, waste sorting, and upkeep of shared public spaces.

• 📐 Top 5% — Fermat Contest, University of Waterloo (CEMC), Mathematics Competition
• 🥉 3rd Prize — HCMC Informatics Competition, Data Structures & Algorithms

I'm interested in applied AI, LLM systems, machine learning engineering, and software engineering—especially opportunities to turn research ideas into dependable, practical products. I enjoy connecting with engineers, researchers, and recruiters.