CV

Graduate Research Assistant

Northeastern University — Augmented Cognition Lab

Research on multimodal ML, video understanding, and human motion estimation, under the supervision of Prof. Sarah Ostadabbas.

• Built AdSelect, a 4,800-pair long-short advertisement dataset mined from ~4M YouTube videos across 17 industries, with a stratified 800-pair benchmark; first paired ad resource large enough to train rather than only evaluate MLLMs on shot selection.
• Formulated ad clipping as a set-prediction problem and developed AdCraft, a LoRA fine-tuning method (Qwen3-VL-8B, rank-512) with complement-shot supervision that injects explicit negative signal to curb over-selection.
• Established a supervised fine-tuning spine (LoRA SFT, rank-512) reaching 0.756 precision / 0.667 IoU, then extended it with complement-shot supervision to 0.771 precision / 0.688 IoU with ~1s duration error, identifying LoRA rank as the primary lever on precision and IoU.
• Surpassed every zero-shot open/frontier model (incl. GPT-5.6-sol) and prior supervised method on the primary metrics.
• Evaluated SFT and RL-based fine-tuning (GRPO via TRL); found greedy selection with complement supervision outperformed RL optimization.
• Ran a systematic zero-shot evaluation of 15+ MLLMs (7B–72B, open, audio-visual, and frontier), showing recall-driven metrics mask over-selection and that scale alone does not solve the task.
• Engineered the full data pipeline: shot boundary detection, Swin3D embeddings, DBSCAN clustering, and cross-duration shot matching.
• First-author paper submitted to AAAI 2027; co-authored companion benchmark (AdShot) submitted to NeurIPS 2026 Datasets & Benchmarks track.
• Extending GVHMR to multi-person scenarios using scene detection, Visual Odometry patching, SMPL mesh overlay rendering, and YOLO tracking on Northeastern Discovery cluster (SLURM).
• Supporting IARPA MOVES pre-proposal at the Augmented Cognition Lab.

Research Engineer

Dashtoon

• Trained and shipped Flux LoRA adapters for 20+ characters across 8 animation styles; validated at 2048 resolution with FP16, batching, and caching optimizations.
• Integrated Hidream-l1 (MoE) into animated character pipeline; improved prompt alignment and output smoothness at 2048 resolution.
• Curated 10,000+ image Bollywood dataset from 100+ movies via automated extraction pipeline to reduce regional bias in generation.
• Evaluated Google Veo 2 (early access) for character animation; partnered with GCP team to integrate workflow into Frameo AI platform.

Project Associate

Indian Institute of Science (IISc) — Vision and AI Lab

• Trained DepthDiT, a Diffusion Transformer for monocular depth estimation, adapting PixArt-Alpha and PixArt-Sigma backbones; best checkpoint AbsRel 0.107, δ1 0.88 at 40k steps.
• Built automated pipeline to curate and QA 30,000 MatrixCity images for depth model training.
• Designed masked diffusion loss for reflective surfaces; implemented DreamBooth and RealFill-inspired inpainting workflows.

Research Collaborator

Carnegie Mellon University (Collaboration)

• Evaluated transfer learning for 3D cryo-ET classification using MedicalNet and ModelNet40-pretrained models.
• Analyzed pretrained representations with t-SNE to assess cross-domain transferability to cryo-ET.
• Converted 3D cryo-ET volumes into point-cloud representations for comparison against voxel-based pipelines.

Senior Research Engineer (Founding Member)

Menmitsu Private Limited

• ITC Freight Damage Mitigation: deployed YOLO-based truck-loading monitoring achieving 70% rough-handling detection accuracy with <5% false positives; led integration of 3 intelligent modules operating in near real-time.
• Smartivity Defect Detection: achieved 97% accuracy on missing-part detection, reducing inspection time from 10 to 2 seconds (5x speedup); built millimeter-precision homography-based deviation detection system.

Data Scientist

EZ Works

• Finetuned Neural Machine Translation models (Arabic–English) exceeding 50 BLEU score; deployed with ONNX-optimized beam search for faster inference.
• Built BERT-based translation memory with custom batch cosine similarity achieving 10x retrieval speedup.
• Implemented LCNN for wireframe and tabular detection in PPT; built document segmentation for image-based documents.
• Built EZ Flip chart/diagram conversion modules with 100% error-free output at 3x speed improvement.

Machine Learning Engineer (Intern)

CyberCure Technologies

• Built LSTM-based demand forecasting for 200+ products across state/territory/beat/shop-owner hierarchy; 90% accuracy on frequently selling products.
• Added convolutional models to mitigate exogenous correlation, improving accuracy and reducing training/inference time by 2x.
• Designed and deployed analytics dashboard for predictions and demand analysis.

Python Developer (Intern)

Arcturus Business Solutions

• Upgraded GUI for Intelligent Surveillance Solution using Python PyQt5 and SQL.
• Added new features and incorporated statistical dashboard into existing system.

Intern

Central Electronics Limited

• Quality check in System Production Department; training in Digital Axle Counter for Indian Railways, Solar Panels, and Microwave Electronics.

Master of Science

Northeastern University

Artificial Intelligence

• Khoury College of Computer Sciences
• Affiliated with Augmented Cognition Lab, advised by Prof. Sarah Ostadabbas

Bachelor of Technology

JSS Academy of Technical Education, Noida

Electronics and Communication Engineering

• Creative Head at Quanta JSS; Member of EDC

Color Masking Method for Variable Luminosity in Videos with Application in Lane Detection Systems

Proceedings of International Conference on Machine Intelligence and Data Science Applications, Springer

A color thresholding method independent of lighting conditions using a two-layered process: luminosity filter followed by variable color threshold. Applied to lane detection for robust performance across weather and time-of-day variations.