Apr 2021 – Sep 2023 · 2 yrs 5 mos · India

• FashionAI
• Built a custom diffusion-based generative model tailored for the fashion domain with fine-grained control over design attributes.
• Curated and engineered fashion-specific datasets to teach the model nuances such as silhouette, fabric texture, patterns, drape, and color harmony.
• Implemented attribute-controlled generation pipelines, enabling precise manipulation of design parameters (neckline type, sleeve length, fit, fabric, color).
• Developed an inference system to generate unique, seller-ready fashion designs, unlocking a new revenue channel for showcasing auto-generated concepts.
• Optimized training using LoRA, DreamBooth, and ControlNet-style conditioning, improving controllability and reducing training cost.
• Competitive Intelligence Platform (CIP)
• Designed and built an end-to-end Competitive Intelligence Platform to benchmark Flipkart against competitors like Amazon across multiple business units.
• Processed data from five major social media platforms to build a unified customer-experience dataset.
• Built NLP pipelines to identify pre-defined customer experience themes (L1 nodes) with associated sentiment using classification + sentiment models.
• Developed a two-stage pain-point discovery system using unsupervised clustering to extract L2 topics, followed by few-shot learning (100 examples) for refinement.
• Implemented a BU-tagging engine to map each feedback datapoint to relevant business units based on product mentions.
• Generated aspect × BU × organization-level performance scores, enabling CXOs to deeply analyze customer experience across business lines.
• Built predictive models using classical ML techniques to forecast:
• Next-week NPS with ~1% MAPE
• Market Share for 8 BUs with 1.5% MAE using ensemble linear regression
• Delivered insights used by CXOs for strategy, marketplace health monitoring, and competitive benchmarking.

Jan 2020 – Mar 2021 · 1 yr 2 mos · India

Jul 2019 – Dec 2019 · 5 mos · India

• This project aimed to recognize characters, emojis, and hand gestures made by moving fingers in the air without wearing sensors.
• 1. Developed and trained a MobileNet version of DenseASPP from scratch on four datasets, achieving an impressive 98% accuracy in hand segmentation using the curriculum learning strategy.
• 2. Additionally, designed an encoder-decoder architecture using a unique combination of 2D and 3D Dense convolution, achieving a validation accuracy of 90% in real-time hand gesture detection.
• This work has promising implications for areas such as human-computer interaction and virtual reality and demonstrates progress towards more advanced hand gesture recognition systems.

Jun 2018 – Jul 2018 · 1 mo

• The project involved the use of CGAN architecture (Pix2Pix) to generate photorealistic composite images with improved aesthetics. A CNN was trained in an unsupervised setting to learn the perception of the visual realism of the composites. Additionally, curated a dataset for image compositing using COCO data, and techniques from image processing, including color transfer using a learnable transfer function, were applied to enhance the dataset.
• Overall, this project demonstrates progress in generating visually appealing composites and showcases the potential for further advancements in this field.