Kai Tarafdar

• Led and orchestrated end-to-end data pipelines, from data extraction and processing to deployment and visualization, using Python, Cloud Storage (Azure, GCS), and Databases.
• Managed and optimized data workflows across multiple pipelines with complex interdependencies, ensuring seamless integration and automation of data processing tasks.
• Developed and maintained taxonomies and ontologies, leveraging linguistic insights and NLP techniques (including BERT, LSTM, and LLM models) to improve data categorization and semantic understanding.
• Designed and implemented data visualization backends for real-time analytics dashboards using Python, enhancing decision-making through actionable insights.
• Guided juniors, providing mentorship on best practices.
• Collaborated with cross-functional teams to ensure that data workflows and analytics solutions met business needs and supported data-driven decision-making.

Develop in house LLM Chat application.

Improve NLP based matching engine. Expand product to multilingual suites.

Improved classification models through data optimization.

• Designed, developed, and implemented accent classification model.
• Performed enhanced EDA of large speech datasets to identify key differentiable speech features.

• Created novel algorithm to generate standardized and clean lexicon from open source lexical resources.
• Composed text normalization and feature extraction library for text-to-speech pipeline.
• Optimized text-to-speech backend to reduce inference time.

• Developed end-to-end NLP and STT projects to be used for production or academic purposes.
• Compiled a retrieval based chatbot using only a small amount of chat transcripts through finetuning on LLMs.
• Compiled a speech to text model achieving SOTA WER with small amount of speech data by leveraging transfer learning method.
• Documented project in technical manuals to be used by clients in future projects.
• Created audio/visual presentation and guides to accompany codebase and documentation.

• Lead interdisciplinary team of linguists and NLP engineers in the creation, expansion, and maintenance of core NLP products.
• Created tools to assist in rapidly expanding NLP corpora, to be used by technical and non technical users.
• Used both classical, recent, and novel methods to process, extract, and classify language features for machine learning.
• Collaborated in development and optimization of end to end custom speech synthesis model.

• Played key development role in very early R&D phase of core new feature of Apple Podcasts: Real-time speech-to-text transcription and Information retrieval.

• Improved and expanded core NLP resources and tools for English locales used with Siri personal assistant.