Senior Data ArchitectVerified as a real position

What you’ll do

• 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 agentic reasoning evaluation — across internal annotators and external vendors.
• Architect the data flywheel: the closed-loop system where real customer conversations feed back into training data collection, curation, annotation, model retraining, and evaluation.
• Own and maintain data pipelines and infrastructure spanning Snowflake, AWS S3, ETL/ELT pipelines (Airflow), and integration with ML training workflows on AWS SageMaker.
• Work directly with LLM, NLU, and Agentic systems teams to understand training data requirements — what conversational patterns improve zero-shot routing accuracy, what dialog structures train better task planners, what edge cases stress-test agentic reasoning — and translate these into concrete dataset specifications and pipeline configurations.
• Define and maintain the data architecture for Omilia's Training Environment: schema design, data flow patterns from production (OCP) to centralized training infrastructure, storage strategy (Snowflake + S3), cross-pipeline consistency, and clear auditable data lineage, including anonymization requirements as part of the compliance layer.
• Design data quality frameworks that directly improve model outcomes: content-based deduplication, diversity-maximizing sampling, confidence-based filtering using NLU scores and behavioral signals, and dedicated NLU improvement corpus extraction from low-confidence and no-match production data.
• Define annotation requirements for ML model development — intent labeling guidelines, entity tagging schemas, dialog act classification, task completion scoring, and reasoning quality assessment — and design annotation workflows that produce consistent, high-quality labels at scale; evaluate and manage external data annotation vendors.
• Build and maintain the data catalog that enables cross-team dataset discovery: document dataset contents, schemas, lineage, quality metrics, intended use cases, and known limitations; define the taxonomy for organizing training datasets across model types (LLM, S2S, NLU, ASR, TTS, agentic).
• Architect the closed-loop data flywheel: production conversations → data selection → anonymization → curation → annotation → model training → evaluation → safe redeployment → back to production; define feedback mechanisms that route model failure cases into targeted training data collection.
• Identify gaps in production training data and define requirements for external data acquisition (public datasets, synthetic data generation, vendor-sourced corpora); design data augmentation strategies for underrepresented languages, domains, or conversational patterns.
• Work closely with LLM/NLU/S2S/ASR/TTS/VB Tech Leads and Senior Engineers to align data architecture with model training requirements; collaborate with Platform Engineering, Security & Compliance, and Product Management stakeholders.
• Maintain comprehensive documentation of data architecture, dataset specifications, pipeline configurations, and data catalog; produce data architecture RFCs for significant changes and share best practices with ML teams.