AI Systems in the Wild
Hemoeco & SupervisorAI
Built chatbots and voicebots, hit infra and reliability limits. Learned that control over state/cache is non-negotiable.
Key Results
Context & Problem
Hemoeco
A construction equipment leasing company wanting to implement a chatbot for preparing quotes and a forecasting model for demand planning.
SupervisorAI
A startup focused on extracting sales insights using AI-generated feedback. I built demos and production systems.
The Critical Inflection Point
Voicebots behaved non-deterministically:
- Pronunciation issues
- Call loops
- Inconsistent behavior across runs
Root cause: State and cache management was abstracted away. I had no control over it.
I learned that AI systems fail silently when state, caching, and infra control are abstracted away.
Rebuilding From First Principles
After leaving, I recreated the entire stack myself:
- Hosted LLMs locally
- Integrated with n8n
- Rebuilt chatbot and voice workflows
- Removed dependency on unstable APIs
Key Learning
Execution speed without architecture eventually destroys momentum, but architecture without ruthless prioritization also paralyzes delivery.
Key Learnings
- This revealed that pricing models shape project success as much as code—hourly billing created misaligned incentives
- This forced me to accept: AI systems demand tight feedback loops; you can't debug what you can't observe
- This taught me that reliability requires control over state and infra—third-party abstractions hide critical failure modes