gyanendralucky

1. Engineered and deployed state-of-the-art Generative AI models, including LLMs, GPT, and LLAMA, to optimize business processes within the insurance sector, resulting in achieving a 10x improvement in customer support efficiency through AI-driven query responses powered by advanced NLP techniques. 2. Implementing AI algorithms for real-time extraction of invoice line items, parsing police reports, and document verification, reducing claims processing times through advanced data analysis and machine learning. 3. Led the development of a sophisticated chatbot creation product, leveraging GPT, vector databases, Langchain, and fine-tuned custom LLMs, which transformed the chatbot landscape from a technical perspective by enabling businesses to create custom chatbots with a single click. Skills: AI/ML, Data Science, Python, AWS, Javascript

1. Orchestrated a comprehensive revamp of the data processing pipeline, harnessing cutting-edge Stable Diffusion and GFPGAN models to refine image synthesis from 10-12 source images. This optimization resulted in an impressive 30% reduction in processing time, vastly improving operational efficiency. This enhancement enabled Aragon.ai to deliver professional-grade photos promptly to clients, enhancing customer satisfaction and streamlining operations. 2. Elevated prompt quality through the strategic implementation of advanced techniques, including the integration of the Codeformer model and automatic eye color change algorithms. These innovations contributed to a remarkable 20% increase in customer satisfaction scores, while also bolstering post-processing methodologies. As a result, Aragon.ai achieved a substantial 25% growth in monthly recurring revenue, cementing its position as an industry leader in AI-powered image generation. Skills: AI/ML, Data Science, Python, AWS

• Implemented active learning to reduce the amount of human annotation required for model training, achieving a 50% reduction compared to random sampling. Demonstrated the effectiveness of these methods using ChatGPT and LLM models for sentiment classification in regional languages. Leveraged zero-shot classification to achieve high accuracy on classes that were not part of the training set, achieving up to 70% accuracy on a set of test classes that were not used in training. • Utilized data augmentation techniques to increase the size and diversity of training data, leading to a 20% increase in data size and a 5% improvement in model accuracy compared to models trained without augmentation. Skills: AI/ML, Data Science, Python, Google Cloud

Developed a few-shot speech cloning model for low-resource languages using active learning methods. Proposed a novel approach to speaker diarization that incorporates the prowess of LLMs to exploit contextual cues in human dialogues. Developed a state-of-the-art audio smoothing algorithm which solves translated audio jarring speedups with the help of optimization and a linear increase in speedup. Skills: AI/ML, Data Science, MySQL, AWS, Google Cloud, Pytest

1. Developed and implemented of an Optimized Text-to-Speech (TTS) model pipeline with multimodal capabilities, resulting in a 30% reduction in production time and a 20% increase in naturalness and intelligibility of synthesized speech, enhancing customer satisfaction and accelerating product deployment. 2. Worked with Phase Modeling in audio source separation through Generative Adversarial Networks (GANs) and Diffuser models, achieving a 25% improvement in audio quality. Skills: AI/ML, Data Science, MySQL, Python

Achieved 95% accuracy in lip synchronization by developing advanced techniques to align lip movements with various audio tracks in human faces. Enhanced talking head video generation from a single image, resulting in a 40% improvement in visual and auditory coherence, delivering high-quality, immersive digital experiences. Skills: AI/ML, Data Science, Python, Google Cloud

• Created an end-to-end service for menu digitization that utilized computer vision algorithms for text detection and recognition to convert paper-based menus into digital format, achieved a 90% accuracy in text detection and recognition, reducing the time required for manual data entry by 80%. • Proposed a two-phase active learning framework to efficiently train a downstream model with limited annotated data by dividing the problem into coarse and fine-grained learning, achieving a more balanced distribution of labeled data for menu data multi classification problem which reduced 30% in annotation Skills: AI/ML, Data Science, Python, Google Cloud

1. Conducted research on the impact of noise in the training corpus on the performance of statistical machine translation systems, specifically examining noise introduced during automatic extraction of parallel corpora from comparable corpora. Results were used to develop strategies to improve system robustness and performance. 2. Designed an online game for natural language processing data acquisition, aimed at collecting high-quality data to improve the accuracy of NLP systems. Successfully launched the game, resulting in the acquisition of a large amount of data in a short amount of time. Skills: AI/ML, Data Science, Python, Google Cloud

Introduce MAViC which leverages our proposed Multimodal Semantics Aware Sequential Entropy (M-SASE) based acquisition function to address the challenges of active learning approaches for video captioning. Our approach integrates semantic similarity and uncertainty of both visual and language dimensions in the acquisition function. Our detailed experiments empirically demonstrate the efficacy of M-SASE for active learning for video captioning and improve on the baselines by a large margin. Skills: AI/ML, Data Science, MySQL, Python, Google Cloud

• Led the design and implementation of edge computing solutions for end users, utilizing Federated Learning in ML for real-time training and inference. • Developed and optimized synchronous training of models using FedAvg algorithm, resulting in significant improvements in model accuracy and training efficiency. Skills: AI/ML, Data Science, Python, Google Cloud

• Utilized probabilistic modeling and Gradient Boosting to reduce credit card settlement failure rates by %, resulting in substantial cost savings and improved financial stability. • Developed a data-driven user affinity model, boosting conversion rates by 10% at checkout, reducing payment failures by 15% during outages, and driving a 7% revenue increase through targeted promotions. • Developed Face Swapper, where user can swap there faces with other images using GFPGAN, RetinaNet, COdeformat, Parsenet Models. Skills: AI/ML, Data Science, MySQL, Python

• Worked on Dynamic Pricing problem where we have to predict what will be the suitable price which should be a sweet point for customer and the service provider, using previous data history of user and location. This model directly impacted on business and increase 10% user. • Worked on Conversion Model where we have to predict the conversion rate of an user to finish the ride. We used the past data of user and geohash features and ride feature to build this model. This reduce the loss by over 3% Skills: AI/ML, Data Science, Python, AWS, Google Cloud, Pytest

1. Worked on Active Learning NLP Imbalance Classification Probelm, where we introduce automatic hardmining which gave us the state of the art performance. Published Internal AMLC(Amazon Machine Learning Conference) paper on the subject of Low Resource Hardmining via Active Learning. 2. Productionized Tobbaco Ads Detection Model, which increase in Precision and Recall of the System by big margin, i.e. 28% Precision Increment and 15% Recall. 1. Worked on Controlled, Optimized, Bidirectional Auto-Regressive Transformer for Ad Headline Generation. Implement multiple strong baselines and show that our method is effectively able to allow control over the length of the generated headline and yield the highest CTR. 2. Demonstrate a 25.82% increment in Rouge-L and a 5.82% improvement in estimated CTR over previously published strong ad headline generation baseline along with a 14.12% improvement in estimated CTR compared to human-written headlines.

1. Developed Amount Detection and Amount type classification from multilingual Indian speech text. Used Question Answering transformer based model with NER(Named Entity Recognition) and we achieve above 99% Accuracy with post-process Technique. 2. Leveraging in house data for predicting location of OkC merchants with high accuracy. Model leverages NEO4J graph DB that we had set up for OkC network. Therein, prediction for each merchant is the most occurring pin code in graph traversal up to a predefined number of hops (set at 5 because of significant decay post that). Location coverage has increased from 11% to 54.7% for the 26.5M merchants registered with us. Skills: AI/ML, Data Science, MySQL, Python, AWS, Google Cloud, Pytest

- Worked on Homepage personalization for 100M+ meesho users - System design, Engineering to serve personalized recommendation at scale - Homepage Personalization, homepage layout and tiles at Meesho are personalised based on user's interest. I work on improving Homepage relevance using ML disciplines like Ranking, Representation learning and Contextual Bandits. Skills: AI/ML