SELECTED WORK · AI PRODUCT / SYSTEM DESIGN
Harmonee
An AI-powered customer engagement system for automotive retail — designed to turn messy CRM history into grounded customer communication by separating what is known from what is merely possible.
Public-safe interactive demo · Fictional customer and CRM data only.
The challenge wasn’t getting AI to write a better message. It was building a system that understood what it actually knew before it wrote anything at all.
Harmonee began as a practical way to help dealership teams create stronger customer communication. But the deeper I went, the more obvious the real problem became: good AI output depends on far more than the model generating the words. CRM history had to become usable evidence. Customer and vehicle context had to stay accurate. Business rules needed to be enforced. Unsupported assumptions had to be caught.
That led to an evidence-driven architecture where Harmonee understands first, protects the integrity of the record, decides what the evidence actually supports, generates the communication, and validates the result before it reaches the user. Missing evidence stays unknown instead of becoming permission to guess. And when something goes wrong, the system makes it easier to tell whether the problem came from the model, the prompt, the data, the workflow, or the system around it.
THE PRINCIPLE
“Beta changes the model, not Harmonee.”
Regular is the stable production experience. Beta allows authorized users to run the same Harmonee workflow through a different AI model. The evidence, prompts, business rules, output structure, validators, history, and telemetry remain the same — so when results change, there is something meaningful to compare.
Evidence Before Generation
Harmonee starts with what the CRM actually says. Customer history, vehicle context, prior communication, and known facts are separated from assumptions before generation begins. Evidence earns specificity; missing evidence requires neutrality. The AI should never sound more certain than the record allows.
Model Lab
The Model Lab turns model testing into controlled research. The same real-world case can be rerun across models while keeping Harmonee’s evidence, business rules, prompts, validation, History, and telemetry consistent — making it possible to distinguish model behavior from problems elsewhere in the system.
Learning From Real Use
Harmonee captures validation outcomes, regeneration, latency, model behavior, and other signals so future decisions can be based on observed performance rather than preference. The goal is continuous improvement without sacrificing trust or stability.