For developers
Building AI that stays trustworthy outside the demo too
Examples, documentation and lessons from actually building AI systems — including the places where the happy path runs out.
What this touches on
The parts where the happy path runs out
MCP and integrations
Connecting a model to a real system: what's right to expose as a tool, what to keep out of its reach, and what an MCP server that's actually maintainable looks like.
Agents and tool calling
When an agent beats a fixed pipeline, and what happens when it errs mid-sequence of tool calls.
Permissions and approval gates
The split between tools that read and tools that write, and exactly where a human approval is needed before an action.
Observability
What you even measure in a system whose output isn't deterministic, and how to catch a regression before your users do.
Cost and latency
Where tokens get burned without anyone noticing, and what that does to response time once there are several calls in a chain.
Production
Rate limits, retries, timeouts and failure states — the part where most demos' happy path runs out.
I built it myself
Code that runs, not a slide
App · Active
Essential
There is more good content than anyone can filter. Essential does the filtering for your role and the tools you follow, and learns from one-tap feedback on every pick.
- Supabase
- Claude
- Brevo
Skill · Active
Eligibility Check for Computing, Digitization, and AI Grant (CEO Directive 4.56)
Businesses do not know if they are eligible for the grant, if their planned expense is recognized, and exactly what needs to be submitted — and the details change between cycles.
- Claude Skills
- SKILL.md
- Browser Tool
- gov.il
Technical articles
What I learned from building
The first technical article is on its way.
Have a technical project worth talking about?
Collaborations, architecture advice or a review of an AI system that's already running — easiest to just write.