The Chatbot Looked Right. The Workflow Told the Truth.
Why enterprise AI should be evaluated by workflow completion, not just answer quality.
Latest
The Chatbot Looked Right. The Workflow Told the Truth.
Why enterprise AI should be evaluated by workflow completion, not just answer quality.
Research desk
Latest analysis
The ATM Agent Boundary: Where AI Should Help and What It Must Never Touch
A practical architecture for LangGraph-orchestrated ATM operations that improves forecasting, triage, fraud detection, evidence assembly, and compliance without putting AI in the cash command path.
Why My AI Service Advisor Forgot Everything
A practical memory strategy for LangGraph workflows, using an automotive service advisor to explain conversation context, execution state, persistent knowledge, and memory governance.
From Automotive Diagnostics to Banking Complaints: What I Learned Building a Local LLM Evaluation Platform
A second domain experiment showing why workflow behavior, retrieval shape, and reusable evaluation architecture matter more than simply adding more context.
From Baseline to RAG to Fine-Tuning: What Actually Works for Automotive AI
A practical experiment comparing baseline, RAG, fine-tuning, and combined approaches for automotive warranty and service diagnostics.
No posts match that filter yet.
About the work
Practical AI systems, tested against real workflow behavior.
These notes focus on evaluation design, retrieval shape, fine-tuning behavior, and the architecture choices that decide whether a model behaves like a workflow or just a helpful assistant.