October2021- Present

Machine Learning Engineer

• Experimented with finetuning of Large Language Model (Mistral) using conversational data sourced from diverse dating apps, to provide context-based advice emulating the role of a personalized dating coach. Implemented advanced techniques including information extraction from screenshot images, to compile and parse data in both instruction-based and ChatML formats essential for training a model.
• Worked on a propensity modeling project for identifying potential customers for all-flash storage solutions for the next 3 to 6-month timeframe using telemetric data. Employed data science skills to clean and simplify data, reducing over 400 features to a concise 250 features, through hypothesis testing and feature selection techniques. Conducted feature engineering to derive new features correlated with the target feature. Trained tree-based models like Random Forest, Decision Trees, and XGboost and achieved an impressive 90% accuracy rate on the test data set. Presented stakeholders with feature importance scores and identified the key features influencing customer propensity.
• Implemented an autoencoder-based fraud detection system utilizing mean squared error(reconstruction loss) as a performance metric, which rapidly decreased once the transactions were confirmed as normal, ensuring precise identification and prevention of suspicious transactions. Developed a 3D Tensorboard Projector Plot using AWS EC2 as a visualization and analysis tool for fraud experts, to help them identify suspicious transactions and prevent fraud.
• Developed a job search engine for an HR consulting firm, incorporating a custom NER model for extracting keywords and named entities, which were subsequently processed by a rule-based data retrieval engine connected to a MongoDB database for efficient data retrieval. Additionally, a query recommender system was implemented to enhance the user’s job search experience by providing intelligent query recommendations using semantic search and FAISS in cases where data was not found in the database.
• Conducted data-driven marketing initiatives by performing exploratory data analysis, implementing KMeans Clustering for customer segmentation, and applying a KNN approach to identify potential customers with similar traits, leading to targeted marketing strategies within a challenging 3-week timeframe and securing a 6-month project extension post a successful proof-of-concept demonstration.
• Designed a powerful information extraction system for invoices and receipts, merging rule-based and machine-learning techniques. Leveraged rules for precise item extraction from vendor documents, and applied Camelot for tabular data extraction, ensuring data accuracy. Integrated this system into an API for convenient access to consolidated item and cost details.
• Researched various time series forecasting models, including the ARIMA model, and SARIMA model as well as modern approaches such as RNN, with a focus on employing XGBoost to improve forecasts through feature engineering in challenging time series data for demand forecasting.
• Implemented concurrent programming (multithreading) techniques to dramatically accelerate data fetching pipelines, achieving a remarkable 15-fold increase in speed. Successfully reduced the time required to fetch one million records from 11 days to just 16 hours. Additionally, implemented a scheduled cron job to enhance the token updation process, ensuring seamless and secure data access by regularly updating access tokens and refresh tokens every hour.

Company Name: FuseMachines