Apurva Bhargava

Experience

• 

  Malaviya National Institute of Technology Jaipur

  May 2018 - Jul 2018

  Research Intern (Machine Learning)

  • Worked individually on single and multi channel static convolutional neural network models and input embeddings for multi-class text categorization and clustering using TensorFlow.• Datasets used include AG’s News, TREC questions and 20newsgroups.• Assisted with data preparation, model evaluation, research paper summarization for ongoing PhD-scholar research work.

• 

  Celeste Online Solutions Pvt. Ltd.

  Jan 2019 - Aug 2019

  Data Science Intern

  • Built face pose classifier in Spark MLlib for exam proctoring, trained on 3GB data using VGGFace for transfer learning.• Detected phishing URLs and malware domains using character-level LSTM networks implemented in Keras.• Collected clients’ website usage statistics, identified KPIs and performed A/B testing to evaluate design choices.

• 

  NYU Center for Data Science

  Jun 2020 - Jul 2021

  • Co-instructed and graded for the Data Science for Everyone course, which is the flagship course of the NYU Center for Data Science, and the first course in the sequences for both the data science major and minor.• Contributed to creating course material (videos and jupyter notebooks) for applied machine learning using Python. • Taught laboratory section (bias in machine learning systems, data profiling, data scrubbing, privacy-preserving data synthesis, machine learning model interpretability) using Python for ‘Responsible Data Science’ course (DS-GA 1017).• Actively contributed to creating course material and testing AI fairness and ML model interpretation frameworks (IBM aif360, fairlearn, mlinspect, SHAP, LIME).• Provided one on one mentoring for assignments, laboratory work and projects. • Taught laboratory and implementation (numpy, pandas, scikit-learn, sampling, significance tests, bias/variance, regularization, logistic regression, decision trees, SVMs, neural networks, bagging, gradient boosting, spam detection, image classification) using Python for 'Introduction to Data Science' course (DS-GA 1001).• Provided one on one mentoring for assignments, laboratory work and projects. • Explored effectiveness of SNIP-technique for pruning CNN architectures (LeNet, VGG-D, AlexNet) in improving efficiency of one-shot and few-shot learning.• Discovered the effectiveness of fine-tuning pruned subnetworks for each task separately in mitigating catastrophic forgetting in multi-task learning. The pruned neurons were masked during fine-tuning.

  • Adjunct Instructor

    May 2021 - Jul 2021

  • Graduate Teaching Assistant | Section Leader and Grader

    Jan 2021 - May 2021

  • Graduate Teaching Assistant | Section Leader

    Sept 2020 - Dec 2020

  • Graduate Researcher

    Jun 2020 - Sept 2020
• 

  New York University

  Jun 2020 - Aug 2021

  (Data Science) Assistant Research Scientist | Graduate Research Assistant

  • Harmonized and combined gameplay and participant survey data from 7 different behavioral studies conducted on-site with farmers in 7 countries as coordination games where coordinated, environment-friendly decisions lead to higher scores– incentivized by actual money.• Studied the effectiveness of PES (Payments for Ecosystem Services) at a very low cost by modeling agricultural and environmental response on subsidies and subjective factors and then performing regression analysis and ANOVA.• Predicted agricultural and environmental outcomes using ensemble models (Random Forest, AdaBoost, XGBoost, LightGBM) with over 92% accuracy in Python; explained causative factors using SHAP explanations and visualizations.• Adapted and tuned Transformers-BART for abstractive summarization of PES research papers (utilized NLP, PyTorch).• Built an Android application using ODK-X for implementing Nash equilibrium games and collecting gameplay data in the background. Show less

• 

  International Flavors & Fragrances

  Jun 2021 - Oct 2021

  Data Science Intern (Finance and Logistics)

  • Performed acquisition, cleaning, and integration of logistics data from multiple databases, spanning 10M data points.• Engineered novel features and built deep learning models in Python for logistics optimization- predicting the busiest routes, estimated times, opportunities to consolidate shipments or change mode from air to ocean, etc.• Forecasted requirements and unit prices for over 10,000 materials from 120+ plants using deep time series models.• Built linear and tree-based extreme multi-class classification models (2082 classes) in Python for predicting the business division of a material by extracting named-entity recognition (NER)-based and text-based features from its description.• Built dashboards for interactive visualizations of models and data at different granularities using Plotly and Power BI.• Presented the models to stakeholders for approval. Show less

• 

  Informed.IQ

  Oct 2021 - now

  Machine Learning / AI Engineer

  Document Field Extraction• Built and deployed deep layout language models for extracting specific fields for 30+ different document types (from loan applications), with an extraction rate and F1 score upwards of 90%.• Collaborated on building automated ML pipelines for document extraction.• Fine-tuned Flan-T5 LLM with reflection tuning and custom loss function for extraction of fields from documents.Fraud or Income Misrepresentation• Built and deployed fraudulent paystub detection model with a 5% fraud flagging rate and negligible false positives (this is one of our top revenue generating products).• Built and deployed near duplicate detection model for detecting bank statement fraud with 3% fraud flagging rate and negligible false positives.• Built graph autoencoder based model utilizing relationships in transaction data for detecting anomalous (fraudulent) bank statements.• Implemented SimHash with Hamming distance and MinHash with LSH indexing in vector databases to explore detection of similar documents.Compliance and Optimization• Built and deployed anonymization model for detecting PII (Personal Identifiable Information), replacing it with random text from same named-entity class with similar font, and automatically blend it in with minimal artifacts.• Successfully executed a POC using a Seq2Seq model for fast sequencing or ordering pages of a document using text in the pages. Show less