Writing
Notes on building cloud, AI, data, and platform systems.
Practical observations from architecture and hands-on engineering: AI systems, cloud platforms, local LLM workflows, reliability, governance, delivery, and value.
How Do You Know the AI Still Works?
A demo can look good and still regress later. Evals turn 'looks good' into something you can repeat, measure, and use as a release decision.
Getting AI Into Production
The model is only one part of production. Identity, networking, cost, observability, rollback, security, and ownership decide whether the system actually ships.
Same AI, Different Constraints
An HR assistant, a fraud system, and a factory-floor model can all be called enterprise AI. The architecture changes because the constraints change.
Does This Really Need AI?
Building an AI demo is easy. Choosing when AI is actually the right tool is harder. I start with the business problem and use the least complicated system that can solve it.
Why Most Enterprise AI POCs Never Reach Production
The model is usually the visible part of an AI POC. Production exposes the less visible work: identity, data access, reliability, cost, governance, delivery, and ownership.
Cloud for Reach. Local for Control.
Local AI is useful as a lab bench and a control boundary. The tradeoff is that once you run the model yourself, memory, quantization, serving, upgrades, and capacity become your...
Building the Products Was the Easier Part
MayaGuru pushed me toward a subscription mindset. ApplyBench taught me almost the opposite — sometimes people just need help at the right moment, not a recurring plan.
Usage Is Not Value
Token usage, tool adoption, and POC counts tell me whether people are using AI. They do not tell me whether the work got better.
What's Actually Been Useful About Hermes for Me
Memory layers, work diaries, background workflows through Discord, and Home Assistant integration — what actually makes an AI assistant stick around instead of being a one-off d...
Why I Moved from OpenClaw to Hermes (for now)
OpenClaw kept breaking — 6 versions in 6 days, one broke Codex auth the day it shipped. So I gave Hermes a real shot, and it felt like a better fit for how I work day to day.