Sagar Chhabriya

Sagar Chhabriya

Multimodal ML & Trustworthy AI

Sukkur IBA University · Sukkur, Pakistan

About

I am a BS Computer Science graduate from Sukkur IBA University (Gold Medalist, CGPA 3.71/4.0). My research focuses on multimodal machine learning, vision–language models, and trustworthy AI—especially medical image–text understanding, retrieval-augmented generation, and rigorous evaluation of language models.

Research Interests

Multimodal Machine Learning · Vision–Language Models · Trustworthy AI · Medical Image and Image–Text Understanding · Retrieval-Augmented Generation · Evaluation of Language Models

Publications

  1. Sagar Chhabriya, Sumra Khan, Inad Alqurashi, Mansoor Ahmed, Rizwan Qureshi. “A Modular Empirical Analysis of RAG Pipelines on WikiQA and SQuAD: The Roles of Chunking, Embeddings, Retrieval, and Model Scale.” Under minor revision, 2026. (First author and corresponding author) [Preprint, Code]
  2. Sagar Chhabriya, H. Dhingana, N. N. Dahar, A. Qadir. “Hirebase: A Cloud-Native Platform for AI-Assisted Recruitment with a Voice Interview Agent.” Bulletin of Toraigyrov University, Physics and Mathematics Series, 2026. [DOI]
  3. Sumra Khan, Sagar Chhabriya, Aizan Zafar, Sheeraz Arif, Amgad Muneer, Anas Zafar, Shaina Raza, Rizwan Qureshi. “Towards Responsible Multimodal Medical Reasoning via Context-Aligned Vision-Language Models.” KnowledgeMR Workshop @ CVPR 2026 (poster); arXiv:2604.08815, 2026. (Second author) [arXiv]
  4. Sagar Chhabriya, Sumra Khan, Rizwan Ahmed Khan, Mansoor Ahmed. “Agentic Retrieval-Augmented Generation: A Compact Review.” Journal of Computers and Intelligent Systems, 3(4):283–296, 2025. [Paper]
  5. Muhammad Rafi, Saif Ur Rehman Awan, Ramsha Jat, Aiman Falak, Fatimah Ansari, Ahmed Raza, Sagar Chhabriya. “FAST-HASOC 2025: Multimodal and Multilingual Approaches for Hate Speech and Offensive Content Detection in Hindi Memes.” FIRE 2025 Working Notes, CEUR-WS Vol. 4173, pp. 625–631, 2025. [PDF]

Research Experience

Context-Aligned Reasoning for Medical Vision-Language Models [arXiv]

Apr 2026

KnowledgeMR @ CVPR 2026 (poster) · Second author

  • Co-authored a framework in which a frozen vision–language model must reconcile radiomic statistics, Grad-CAM activations, and vocabulary-grounded cues before returning a structured chest X-ray answer with evidence, uncertainty, and safety notes—reducing fluent but weakly grounded conclusions.
  • Conducted cross-dataset evaluation on CheXpert; radiomic features achieved AUC 0.71 while text embeddings reached 0.50.
  • Found that image-plus-text conditioning reduced average hallucinated keywords from 1.14 to 0.25; tool-augmented reasoning shortened evidence explanations from 19.4 to 15.3 words without increasing model confidence.
  • Designed the methodology figure, quantitative tables, and conference poster.

A Modular Empirical Analysis of RAG Pipelines on WikiQA and SQuAD [SSRN]

2025–2026

First author and corresponding author · Under minor revision

  • Isolated four design levers—6 chunking strategies, 9 embedding models, dense/BM25/hybrid retrieval, and 7 generators (82M–7B)—across 44 main configurations and 18 ablations on WikiQA and SQuAD.
  • Built a reproducible BEIR-style benchmark showing that corpus construction errors alone can force Exact Match to zero regardless of the generator.
  • Reported retrieval metrics (Hit@k, MRR, answer-in-context) beside answer quality with 95% bootstrap intervals and paired significance tests.
  • Found that the generator dominates (SQuAD EM 0.00 → 0.65 from 82M to 3B instruction-tuned), BM25 trails dense retrieval (Hit@5 0.754 vs 0.902), and chunking has no significant effect.

Sentiment Classification of Hindi–English Memes (HASOC 2025) [FIRE Working Notes]

June 2025 – Aug 2025

NLP Research Assistant, VSL Lab, FAST-NUCES · Karachi, Pakistan

  • Developed a TF-IDF and Random Forest pipeline for ternary sentiment classification of noisy OCR-extracted Hindi–Hinglish meme text (1,133 training samples).
  • Diagnosed weak baselines via manual instance inspection (OCR noise, code-mixing, task-specific patterns) and curated Hindi–Hinglish stopword and offensive-term dictionaries.
  • Improved macro-F1 from ~0.64 to 0.70 on the test set with five-fold stratified cross-validation; co-authored the FIRE 2025 Working Notes paper.

Education

Bachelor of Science in Computer Science

Sept 2022 – Aug 2026

Sukkur IBA University (CGPA: 3.71/4.0) · Sukkur, Pakistan

  • Distinction: Gold Medalist
  • Relevant coursework: Machine Learning, Artificial Intelligence, Computer Vision, Data Science, Probability and Statistics, Linear Algebra, Design and Analysis of Algorithms, Human–Computer Interaction

Teaching & Work Experience

Machine Learning Trainer, atomcamp

Aug 2025 – Present

Islamabad, Pakistan

  • Mentored 80+ learners and guided 30+ students through end-to-end AI/ML projects and deployments.
  • Conduct weekly sessions on ML, DL, NLP, Generative AI, MLOps, Docker, CI/CD, and cloud deployment.
  • Designed capstone projects, coding labs, and technical guides; provide 1:1 mentorship and code reviews.

Programming Fundamentals Course Instructor, Times Tech Computer Center

Feb 2025 – May 2025

Kandhkot, Pakistan

  • Delivered a four-month C++ course to 15+ secondary-school students covering core syntax and structured programming.
  • Designed 25+ hands-on exercises and debugging activities.

Selected Projects

Hire-Base (AI-Assisted HR & Recruitment Platform) [GitHub]

React, FastAPI, PostgreSQL, Supabase, Azure, LiveKit, Groq, Deepgram

  • End-to-end HR platform for onboarding, hiring, assessments, interviews, and recruiter insights.
  • AI voice-interview agent with LiveKit, Groq/Llama 3.1, Deepgram STT/TTS, and VAD; CI/CD via GitHub Actions.

Graph Neural Networks for Aerodynamic Drag Prediction

PyTorch, PyTorch Geometric, GraphSAGE, CFD

  • GNN framework predicting aerodynamic drag from airfoil CFD meshes (nodal velocities + connectivity).
  • GraphSAGE with Top-K pooling; test R² 0.9837, outperforming CNN and shallow regression baselines.

Skills

  • Programming: Python, C++, Java
  • ML & NLP: Scikit-learn, PyTorch, TensorFlow/Keras, NLTK, Hugging Face Transformers, FAISS, LangChain
  • MLOps & Cloud: Docker, GitHub Actions, W&B, GCP Cloud Run, AWS (EC2, S3), Azure Web Apps
  • Web & Tools: FastAPI, Streamlit, Git, GitHub Codespaces

Talks

  • 1st ICCA @ Salim Habib University (May 2025) — Presented: Agentic Retrieval-Augmented Generation: A Compact Review

Awards & Leadership

  • Gold Medalist — BS Computer Science, Sukkur IBA University
  • Hacktoberfest 2025 Super Contributor — Recognized open-source contributor
  • Master Trainee PVE (EU-funded) — Trained 185+ students in violent extremism prevention
  • GitHub Foundations Exam — Certified in Git, GitHub Actions, and project management
  • Aspire Institute Leaders Program 2024 — Completed Cohort 1 leadership training
  • Prime Minister's Laptop Award 2023 — National merit recipient for academic excellence in Computer Science
  • Book Review Competition Winner 2022 — 1st place among 17 sections across 5 SIBAU campuses