From Lab To Life: Practical AI System Evaluation
Talkat Open Source Summit Japan + AI_dev: Open Source GenAI & ML Summit + Automotive Linux Summit 2025 · 2025-12-08..12-10·
summary
Argues that fixed benchmarks like MMLU miss real-world risks and demonstrates a risk‑adjusted, continuous evaluation pipeline that moves from design to production using OpenTelemetry for unified traces, MLflow and Langflow for trajectory auditing, and LLM-as-judge scoring inside RAG and multi‑agent flows. Shows an open‑source Finos case (Finsight) that uses Langraph, metacognitive reasoning, and dynamic guardrails, reports using 17 design patterns and reusing ~200 validated trajectories, and claims the system self‑corrects in production.
AI-written summary · about
chapters (20)
- 0:00AI system evaluation intro
- 1:28Traditional evaluation gaps
- 3:17Chatbots to agentic AI
- 5:10What is agentic AI
- 8:02Benchmarks fall short
- 10:15Theoretical evaluation framework
- 13:24Dynamic evaluation methods
- 16:16Core evaluation strategies
- 18:43Inner loop for developers
- 21:13Auditability and trajectory
- 23:42Custom metrics with LLM judges
- 25:09Debugging with MLflow
- 27:17Production observability
- 30:10Real-world execution paths
- 31:42Bare-metal metrics & human feedback
- 33:52Human roles in evaluation
- 35:19Case study: Finsight agent
- 37:05Metacognitive reasoning
- 39:14Multi-agent architecture
- 40:41Key takeaways
related talks
47mHow to Build Enterprise-Grade AI Agents Using Robust Evaluation
43mEval-Driven Development: Mastering Agentic Tracing and Expert-Aligned Judges With MLf...
01h 28mMastering RAG Evaluation with MLflow & Open Source LLMs
16mTaming Rogue AI Agents with Observability-Driven Evaluation
01h 26mShipping AI That Works: An Evaluation Framework for PMs
26mAI Observability and Evaluation: Ensuring Trustworthy Intelligent Systems
36mAdvanced Evaluation Techniques for AI Agents
33mEvaluating Your AI Agent