TikTok Recommendation team - Push Notification, EDM & Story

• Worked on Hila - a generative AI research assistant that answers questions on proprietary enterprise documentsNamed Entity Recognition: • Reduced Hila’s search latency and cost by training and deploying an NER model• Trained a custom spaCy NER model on financial queries dataset generated by prompting GPT 3.5 turboTopic Modeling:• Built Hila’s multi-document visualization which discovers latent topics using BERTopic.• Used all-mpnet-base-v2 sentence transformer to map textual documents to 768 dimensional vector space• UMAP for dimensionality reduction to 5 and HDBSCAN for clustering documents belonging to same topic. • For each topic, obtained representation using tf-idf and finetuned it further using GPT 3.5 / Llama 2PDF Tabel Extraction:• Built a pipeline to extract tables in Thai Language with the help of Yolo v8s detection model and easy OCRAnti-Hallucination:• Fine-tuned BART model on negation dataset

• Designed, implemented and deployed customer billing and API gateway changes for Azure Premium V2 disk storage which is now generally available (GA); added unit and integration test coverage for these changes• Implemented, deployed and tested incremental snapshot feature changes in gateway of Azure Ultradisk• Added metrics & monitors to track changes in production; gained systems programming experience in C++

Interned in SecDB team in Engineering Division at Goldman Sachs. Implemented database index CRUD operations in a middleware (written in Rust language) that can allow transition from in-house database to Amazon DynamoDB without breaking backward compatibility