ML Intern - Embedded AI Lab, LUMS - Lahore, Pakistan
(2025-05 - 2026-06)
- Designed a training objective combining invariant regularization, contrastive alignment, and MoE architecture to improve robustness on unseen domains, gaining +5-10% accuracy over ERM on PACS and VLCS in DomainBed.
- Trained a lightweight Transformer verifier on a synthetic step-level error dataset to catch and fix LLM reasoning errors without full regeneration, applying localized edits to flagged steps and achieving +6-12% accuracy on structured reasoning benchmarks stabilizing after 2-3 correction steps
ML Intern - Computer Vision and Graphics Lab (CVGL), LUMS - Lahore, Pakistan
(2025-03 - 2025-12)
- Fine-tuned DeepSeek-Coder-7B with LoRA on a 75k synthetic text-to-OpenSCAD corpus to enable non-experts to generate 3D floor plans from natural language, achieving ~74% semantic accuracy and 100% syntax validity, outperforming the Tell2Design baseline on held-out data
- Built a multi-step user feedback layer allowing iterative layout refinement without restarting from scratch, maintaining >70% structural consistency across successive revisions
ML Intern - Intelligent Machines and Sociotechnical Systems Lab, LUMS - Lahore, Pakistan
(2025-08 - 2025-12)
- Augmented Double DQN with a learned dynamics model to reduce Q-value overestimation via short-horizon rollouts during action selection, reaching ~80% of final reward 20-35% faster with 12-25% lower Q-value variance on Pong and Seaquest across 5 random seeds
ML Intern - Center for Speech and Language Technology (CSALT), LUMS - Lahore, Pakistan
(2025-05 - 2025-08)
- Curated and annotated ~2,500 long-form YouTube videos to train a GCViT-based classifier for detecting AI-generated content at scale, reaching ~75.7% accuracy on real-world held-out data
- Extended the baseline with pseudo-labeling, active sample selection, and a multimodal fusion head visual+audio to capture synthetic cues missed by vision, reducing labeled data needs by ~22-30% and improving accuracy by ~4-6%
ML Intern - Human Interaction and Social Experience Lab (CHISEL), LUMS - Lahore, Pakistan
(2025-03 - 2025-08)
- Built an AI-assisted food logging prototype using image recognition and quick-entry shortcuts to address poor calorie tracking support for South Asian diets, cutting manual corrections per meal by ~30% and reducing abandoned logging sessions across 100 users reduced entry time to ~5 s (quick logs ~15 s, custom entries 40–55 s), with ~80% ease-of-use and 85% learnability.
Teaching Assistant - Lahore University of Management Science - Lahore, Pakistan
(2024-01 - 2025-12)
Courses: CS 437 Deep Learning, CS436 Computer Vision, CS331 AI and Machine Learning, CS200 Intro to Programming
- Supported 400+ students across three CS courses by designing labs, quizzes, assignments and tutorials that scaffolded implementation of ML models in Python, OpenCV, and core ML concepts and surfaced common misconceptions for targeted feedback
