Juhi Chandalia

• Maintained quality of model in production by developing metrics and visualizations to facilitate conversations with stakeholders and pinpoint areas for improvement. • Prototyped solutions for new business areas using generative AI models like T5 and vision-based models like Donut. • Led adoption of Airflow with Kubernetes for our model training pipelines and model performance alerting.

• Increased automation rate of the model powering a key product from 60% to 80% by deploying and maintaining a BERT-based LLM (large language model) for text classification.

• Led data visibility at the company to provide insights to business and technical teams. • Prototyped a dashboard in Python using Dash and Plotly using data from a MongoDB. • Trained deep learning models using TensorFlow to determine viability of new business areas. • Wrote ad-hoc scripts for tasks such as filtering data, client requests, interfacing with hardware.

• Created Toximeter, a Chrome extension for identifying toxic comments on Twitter. Used GCP and tensorflow to train a DistilBERT-based model and then deployed the model using a Google Cloud Function. • For a FinTech startup, worked with CEO and business solutions team to create alignment around machine learning approaches. - Worked with finance experts to extract domain knowledge into a set of features appropriate for machine learning. - Used AWS infrastructure and python, PySpark and SparkSQL to do preliminary data visualization and feature extraction in preparation for unsupervised and semi-supervised techniques. - Coordinated internal and external resources to build a new analytics group. • For a marketing startup using beacon technology, guided product thinking around A/B testing and data visualization for the product and customer success teams. • For a large integrated circuit manufacturer, used signal processing and machine learning techniques to pinpoint the key parameter responsible for faulty product lots. Used results to assist the business development team gain consensus towards building a smart facility. • For a medical device nonprofit, developed a novel metric to determine improper running form using accelerometer data and signal processing techniques in Python. • Used Python-based NLP techniques (fuzzy string matching, n-grams) and clustering to normalize a large text database for the RND arm of a large software company.

• Use Deep Learning techniques for computer vision problems in support of Sentient’s retail product. - Use keras and both publicly available pre-trained networks and internally developed networks running on GPUs to produce embedding spaces for customer images. • Manage a 300K project to prototype Sentient’s next product. - Tune the compute resources and optimize performance of the Spark clusters used in the project. - Coordinate internal (DevOps, software architects, machine learning specialists) and external resources (subject matter experts) for the project. - Used AWS infrastructure and python, PySpark and SparkSQL to do data munging, feature engineering, and machine learning.

• Led a project to annotate Affectiva’s full video dataset with emotion content. - Used Amazon Elastic MapReduce to process videos using Affectiva’s proprietary emotion analytics technology. - Crowd sourced data labeling using Amazon Mechanical Turk. - Used python to perform inter-rater agreement and clustering on the resulting data to reveal patterns in the U.S. advertising market. • Researched current thinking about blink rate as an indicator of viewer engagement and translated findings from controlled lab settings into testable hypotheses for Affectiva’s data. - Used python to process raw data (interpolation, detrending, baselining, sigmoid transformation) to create usable signal trace data and individual and aggregate-level blink metrics to investigate blink synchrony and blink suppression engagement measures. • Performed data visualization and created a logistic regression model in python for gaze direction using facial tracking points and head pose data.

• Performed ad hoc machine learning, statistical modeling, and clustering analyses. • Prototyped (in R and SQL), tested, productized (with PL/SQL) and validated multiple predictive models for applications such as predicting customer churn in the mobile space. • Analyzed petabyte-sized datasets (using python) from trillions of records per month, derived statistically valid and actionable insights from the data, and presented findings as compelling business cases for areas such as network abuse, analysis of pricing plans based on actual usage, enhancement of advertising offers, churn reduction, and customer care. • Interfaced with customers in order to design effective solutions and presented recommended actions back to customers.

Completed coursework and qualifying exams for doctoral program.