Projects
Fine-Tuning LLMs for Math Reasoning While Preserving Safety Alignment
Fine-tuned Qwen2.5 models (1.5B & 7B) on the GSM8K dataset using LoRA, improving math accuracy from 38% to 81% while maintaining a strong 88% safety alignment score on the AILuminate Safety Dataset.
Performed ablation studies across 10 different hyperparameter configurations (learning rate, LoRA rank, dropout) to analyze performance trade-offs and mitigate catastrophic forgetting in fine-tuned models.
Post-Training Quantization (PTQ) for Diffusion Transformers
Conducted a head-to-head empirical study, comparing low-bit floating-point (FP) and integer (INT) quantization formats on a PixArt-α (0.6B) diffusion model under similar precision budgets.
Observed that FP quantization yields better visual fidelity at low precision, improving FID from 42.4 to 38.8 under W4A8 compared to INT PTQ with comparable CLIP and ImageReward scores.
Deformable Object Manipulation with Vision-Language-Action Policies
Trained SmolVLA on teleoperated demonstrations using imitation learning for long-horizon control tasks, achieving 69% success rate on garment folding, outperforming Diffusion Policy (41%) and ACT (61%) baselines.
Enhanced model generalization by augmenting training with NVIDIA Cosmos-Transfer synthetic data, exposing the policy to diverse physics-grounded visual scenarios.
Rotation Invariant Multi-Object Detector
Addressed the limitations of traditional object detectors struggling to identify improperly oriented images by integrating eigenvector analysis and a custom decision criterion with a pre-trained YOLOv3 model.
Achieved a 43% accuracy improvement over the ResNet50 model on the Pascal VOC 2012 dataset for images deviated between 90°–270°, without any rotation-augmented training.
Deep Reinforcement Learning for Real-Time Bidding (RTB) in Sponsored Search: A Literature Review
Analyzed Deep Reinforcement Learning (DRL) applications in Real-Time Bidding (RTB), formulating the bidding process as a constrained Markov Decision Process(MDP) to address highly non-stationary market behaviors.
Evaluated Deep Q-Network (DQN) architectures - comparing Control-by-Model state aggregation against Direct-Action continuous adaptation - highlighting strategies that improve long-term budget pacing and achieve up to a 120% ROI increase over traditional ones.