Automatic classification pipeline for supplier compliance documents — 10,000+ documents per day processed in production. Costs divided by 2 to 3 compared to a naive per-call GPT-4 approach, using batching, model selection, and structured output caching.
Automated extraction of structured insights from sales call recordings — summaries, detected objections, competitor mentions, and next steps. Integrated with the existing CRM via webhook. Reduced manual note-taking time by over 60%.
Systematic reduction of LLM inference costs across multiple production pipelines — model routing, prompt compression, response caching, and batching strategies. Achieved a 2–3x cost reduction with no degradation in output quality.
Projects currently in development or recently completed — details to follow.