MSFT AI: LLMs + NLP in M365 Copilot Search
• Engineered an automated DSAT triaging LLM pipeline for Language Understanding, achieving 97% recall and 92% precision across 1,000+ queries.
• Developed a privacy-preserving anonymization algorithm with 99%+ identifier removal accuracy across 50k+ DSAT logs, eliminating PII leakage and enabling compliance-safe model training.
• Scaled DSAT infrastructure to triage 100+ cases weekly, unlocking high-confidence fine-tuning and reducing turnaround time by 5×.
• Built a human-in-the-loop UI and domain-optimized prompts, streamlining triage validation and accelerating model alignment.
• Cut annual DSAT processing costs from ~$430k → ~$4.2k, driving a 100× efficiency gain through automation and workflow redesign.
