Work Experience

To address the issue of unplanned crew absenteeism, I implemented a predictive system that forecasts crew absenteeism on a day-of-operation basis. The system leverages a comprehensive database that includes crew profiles, historical absenteeism records, and roster activity information. Given a published roster period, the system applies machine learning techniques to predict absenteeism probabilities for all crew members scheduled to be on duty for a given day of operation. The system contains features like experiment tracking, model monitoring, and hyperparameter tuning, which can be easily managed and visualized using the MLflow Dashboard. Additionally, the system incorporates LIME (Local Interpretable Model-agnostic Explanations) to provide interpretability and transparency in its explanations. The reports produced by the model allow clients to understand the factors that influence absenteeism, enabling them to leverage these predictors effectively. Further, I worked on analyzing crew roster stability and fairness using ARIMA, graph outlier detection algorithms, and clustering

Through this project, I have fortified my skills in MLOps. I learned how to create a model registry in MLOps and containerize ML models. As new data comes in every day, I also learned some data-versioning techniques. Further, I  have shown tremendous expertise in converting raw data to well-visualized insights. 

I had the privilege of working with the brilliant team at DAI Lab, MIT, led by Kalyan Veeramachaneni, on time series anomaly detection. Orion, a Python library developed by the lab, is designed for unsupervised time series anomaly detection. Due to limited labeled data, most research in this area relies on a few well-known benchmark datasets, such as NASA, OMNI, Yahoo, and Numenta. As part of my thesis, I analyzed these benchmarks and built pipelines that could uncover points of triviality and run-to-failure bias (RTFB). These pipelines, which were published to the Orion library, revealed significant insights, such as 18% triviality in the OMNI dataset and 30% RTFB in the Yahoo dataset. This work sheds light on the limitations of current benchmarks and aims to push the field toward more robust, reliable evaluation methods.

During my time at the DAI Lab, I sharpened my ability to sift through research papers and pinpoint areas where immediate improvements could be made. I developed a solid understanding of various anomaly detection algorithms, ranging from traditional methods like ARIMA and statistical thresholds to advanced deep learning techniques such as autoencoders and LSTMs. Beyond the technical knowledge, collaborating with a skilled team provided me with invaluable exposure, allowing me to refine my ideas and contribute to more impactful research.

Souce Code: https://github.com/MihirT906/Orion/tree/SOL-pipelines

During my time at Raksul, I played a key role in enhancing JOSYS, a SaaS management platform that empowers businesses to integrate, automate, and manage their software applications seamlessly. While many applications offer APIs for integration, some don’t, posing a significant challenge. To tackle this, I developed an RPA (Robotic Process Automation) solution for API-less integration. Using Puppeteer, a web scraping library, I built an automation bot capable of performing critical tasks like signing in, retrieving users, inviting or deleting members, and even deleting accounts. Additionally, I developed a Chrome extension to monitor user activity across select SaaS platforms, providing actionable insights into app usage. This feature enabled businesses to identify underused applications, optimize their software stack, and significantly cut costs. Through these contributions, I helped enable secure integration for 25% of the company’s systems by implementing OpenSSL for authentication, a pivotal effort that directly contributed to the company securing $30 million in funding.

• Conceptualized an interactive story-based Alexa Skill called “Swallowed by the Sea”. The design provided wide exposure to creative writing, conversational AI design, and Voice User Interface (VUI) design.

• Implemented the Alexa skill using node.js and Python and was also introduced to the fundamentals of Natural Language Processing (NLP). 

• Developed a language learning application called Trilingo to revive endangered tribal languages. The learning engine incorporated the Spaced Repetition algorithm (Leitner's) to ensure that the learning was incremental and effective. 

• Implemented the cross-platform app using React Native.