Eduardo Lizana

• • Client engagement: building the agent harness and infrastructure for a mid-market company to become AI-native, including a production checkout-inventory agent that ingests post-stay data and autonomously decides what to charge tenants for and what requires ops action. Stack: Python, LangChain, DeepAgents.

• • KMC RAG — LangGraph router agent over 503-file corpus. Hybrid retrieval (vector + BM25/RRF), contextual enrichment via subagents lifted accuracy 65% → 85%. Used RAGAS evaluation.

• • Designed custom orchestration and context engineering layer from first principles: node-based goal system with directed graph flows, 5 concurrent agents per call each with scoped context. Multi-provider LLM integration (Anthropic Claude, OpenAI, Groq, Cerebras).

• • Built tiered evaluation system across all agents: code-based assertions, LLM-as-judge, pairwise model comparison, and human expert review. Evals drove architecture changes and model selection decisions.

• • Engineered streaming voice pipeline: LLM tokens split at punctuation → TTS per phrase → parallel prefetch → interruption detection at every stage. Evaluated 5 STT vendors hands-on before selecting.

• • Shipped 2,000+ real patient calls across 3 countries.

• • Reduced burn from $750K to $150K/month (80% reduction) while driving 77% revenue growth. Cut cost per delivery 23%.