I led model alignment and evaluation for Russian voice synthesis models at micro1, closing gaps in phonetic grounding and prosodic naturalness by converting human evaluations into engineering-ready data pipelines. Key Architectural Results: • Audited 1,000+ multimodal data samples, cutting output inconsistencies by 18%. • Engineered a structured error taxonomy for phonetic features and pitch contours. • Integrated HITL validation into sprint cycles, reducing model training latency. • Standardized rubrics to optimize production-facing alignment.