RAJ***IAH

Generative AI, Artificial Intelligence (AI), Machine Learning, Natural Language Processing (NLP)

AI+ML+NLP Principal Machine Learning (ML) Engineer and Senior Generative (Gen) AI Architect at GPSUSA.AI

• Location: San Francisco, California, United States
• Duration: 2014-09 to Present
• Description: JOB DUTIES / RESPONSIBILITIES: Building predictive models and state-of-the-art NLP/machine-learning algorithms to interpret, make inferences and offer search and recommendations based on the data's insights. Employing Transformer-based Pre-trained Language models like Bert and GPT-3 for downstream NLP/NLU tasks and bringing them to production. Processing petabytes of unstructured text data for the entity, relationship, and signal extraction. Applying cutting-edge algorithmic techniques and data structures in Machine Learning, Natural Language Processing (NLP), Data Science, and Deep Learning. HANDS-ON PROGRAMMING EXPERTISE: • Customizing ChatGPT, BERT and GPT-4, and other cutting-edge Transformers / Large Language Models (LLMs) for NLP Pre-training, Prompting and Fine-Tuning applications. • Generative AI / Machine Learning: SciKit-Learn, Keras, TensorFlow, Pytorch • Search and Recommendation Systems for User Personalization with collaborative FIltering • Deep Learning Algorithm Used: Word2Vec, RNN, CNN, R-CNN, LSTM, GRU, Sequence-to- • Sequence, Reinforcement Learning • Deep Learning Tools Used: Spacy, Gensim, Apache Spark, TensorFlow, PyTorch • Programming Languages: Python, R, C++ • Computational Linguistics NLP Tools: Stanford Parser, NLTK, OpenNLP, Spacy, Gensim • Information Extraction/Retrieval Tools Used: Stanford POS, and NER (Stanford • CoreNLP), OpenNLP • Information Retrieval / Open Source Search Engines: Solr, Lucene, Elasticsearch • Machine learning Tools: Scikit-Learn, Spark, Pyspark • Machine Learning Algorithms Used: Decision Tree, Naïve Bayes, Random Forest, Gradient Boosting Machine (GBM), XGBoost, SVM • Topic Detection: LSI, LDA, K-means • Semantic Web Technologies: Semantic Ontologies, Knowledge Graphs. Neo4j, Cypher, Sparql • SQL: PostgreSQL, MySQL, MS SQLServer, OracleDB • BioBert for Healthcare and Biomedical Text Analysis: NER and Relationship Extraction, Knowledge Graph from EHR and EMR • MLOps

Principal Machine Learning Engineer and Senior Generative AI Architect at GPSUSA.AI

• Location: Palo Alto, California, United States
• Duration: 2019-07 to 2025-07
• Description: Conversational AI: BERT and GPT-3 Language Transformer models for NLP Applications and Dialogue Systems We use massive deep learning language models like Bert and GPT-3 for language understanding and use them in many NLP applications like Information Retrieval (IR), text classifications, question answering, and other applications. We fine-tune these pre-trained models so that downstream users can create task-specific models with the smaller training dataset. This technique is called transfer learning. We gain context in these models. These models use an attention mechanism, which “pays attention” to the relevant words to contextualize and add a semantic edge to our NLP task-specific models. The Transformers (such as Bert and GPT-3) are highly parallelizable, can train larger models faster, and use contextual clues to fix many ambiguity issues that plague text. • Our NLP applications use Bert as its use of bi-directional learning to gain context of words from both left-to-right and right-to-left gives precise semantic meanings to enable the model to understand how sentences relate to each other. • We used Bert to provide deeper text understanding for IR. Compared to bag-of-words retrieval models, the contextual language model can better leverage language structures, bringing considerable semantic improvements in queries written in natural languages. • We use these models in natural language generation (NLG) to automate communications, report writing, summarizations, question-answering platforms, and search query understanding.

Principal Machine Learning Engineer and Senior ML Systems Architect at GPSUSA.AI

• Location: San Francisco Bay Area
• Duration: 2019-06 to 2025-07
• Description: CONVERSATIONAL AI, and CHATBOT DESIGN/DEVELOPMENT: Used state-of-the-art Pre-trained Language models for NLU/NLP tasks. • Intent Classification: - Used various machine learning techniques like Random Forest, Naïve Bayes, and Decision Tree for intent Classification. • Query Parsing and Entity Recognition: - Used CRF algorithm for parsing natural language query and named entity recognition. Also Integrated Spark-SQL and Apache Hive for handling generated SQL Queries against large database. Solr is used to support query for ad-hoc report. • Conversational Unit: - Integrated Chatbot with Conversational agent -- Rasa-NLU. • Used Language models for natural language generation (NLG). • Speech Recognition: - Trained Kaldi with context free grammar (CFG) and Deep learning methods. Also used HMM’s for speech recognition. Currently using Machine Translation using Attention Based Model for Speech Processing. Deploying LSTM – Based Neural Language Model for speech recognition with Attention. • Gesture Recognition: - Used object detection algorithms like Regional-CNN (Fast R-CNN) to detect fingers and eyes expression and then Bi-LSTM is used to classify gesture sequence. Data Science, NLP, Natural Language Processing, Analytics, Artificial Intelligence, AI, Machine Learning

Principal Machine Learning Engineer, Staff Data Scientist E-Commerce Search and Recommendations at GPSUSA.AI

• Location: Palo Alto, California
• Duration: 2018-06 to 2024-04
• Description: Creating the Next Generation Search Relevance, Recommendations, and Ranking Platform for User Personalization Our Search, Recommendations, and the Ranking platform uses personalized recommendations to help customers find desirable content quickly. Our Search engine is one of the most powerful influencers when consumers consider which brand or product to purchase. It provides personalized service to every client by analyzing customer data and then using it to create accurate, individualized, and customized client profiles. Our Product recommendation software delivers content based on estimates of what the customer wants or needs. Simply put, it intelligently anticipates the customer's intent and then provides a unique personalized recommendation based on observations. The click-through rate of our personalized recommendations is twice as effective as the click-through rate of non-personalized recommendations. Our Ranking algorithms are used most often in situations where the results of a query or request need to be ordered by some criterion. Our Recommendation systems use these algorithms as a final step to sort suggestions by relevance before presenting the results to the user. It helps retailers deliver the right offer at the right time to the right shopper, resulting in a likelihood of conversion and more money spent per transaction. We are witnessing a 35% increased revenue generated by our recommendation engine through natural, upsell, and cross-sell opportunities. A. Recommendations Systems (or Recommender Systems) • Building probabilistic retrieval models, statistical language models, query likelihood retrieval functions for information retrieval, and search engines. • The best solution we have designed is the hybrid system of combining content-based filtering and collaborative-based filtering with the weighted average • We achieved state-of-the-art results with the hybrid approach of Restricted Boltzmann Machines (RBM) and SVD++

Principal Machine Learning Engineer and Senior Staff Data Scientist Ads Ranking at GPSUSA.AI

• Location: Palo Alto, California, United States
• Duration: 2018-01 to 2021-07
• Description: Digital Advertising Platform: Choosing the right ads for the user and the order in which they are displayed greatly affects the probability that a user will see and click on each ad. Also, ads ranking has a strong impact on the revenue received from the ads. So it is important to estimate the CTR of ads in the system accurately. Our machine learning models predict the CTR for new ads effectively. Advertisers care more about ad performance, which drives a high motivation of data science for digital advertising optimization. But consumers value and respond to personalization – keeping it relevant, timely, and contextual – but won't tolerate being bombarded with poorly timed, intrusive, or irrelevant messages. We harness the power of personalization to target individual shoppers with the right message, at the right time, by contextualizing the ads using machine learning at scale. As cookies begin to phase out, we use state-of-the-art contextual models and algorithms to serve relevant ads based on a real-time determination of a webpage's content, thereby enabling the cost-efficient placement of well-tailored assets that attract interest and consideration. • Our contextual intelligence, driven by machine learning and Natural Language Processing, enables the programmatic purchasing of digital advertising based on appropriate categories of relevance. • Contextually driven asset placements inspire a statistically significant increase in purchase intent. • Our contextual targeting uses Natural Language Processing and Computer Vision to better understand unstructured data -- like text, images, and video -- making it a sophisticated way for ad target optimization with relevant, compelling content and other assets placed in front of interested eyeballs. • The state-of-the-art ads ranking system uses a combination of semantic and syntactic features for ad matching and ranking. ORTB: We use Optimal Real-Time Bidding for Display Advertising using Survival models.