Own the Training Environment data architecture end-to-end: dataset design and schema for all ML training pipelines, including dialog corpora for LLM training, conversational steps for NLU models, annotated evaluation sets, and whole-call recordings for speech-to-speech model development.
Define and govern data selection and sampling strategy: establish criteria that determine which production conversations have the highest training value, including diversity-optimized sampling, confidence-based filtering, edge-case prioritization, and deduplication strategies.
Build and maintain the data catalog and dataset discovery infrastructure: enable ML engineers across LLM, NLU, Speech, and Agentic teams to find, understand, and use training data without friction.
Define annotation pipeline architecture: establish requirements for data labeling — intent annotation, entity tagging, dialog act classification, task completion scoring, and age...