Notes from production
What I learn building and running AI and automation systems for clients: project write-ups, costs, tools, and what broke.
Jev Lowered One Workflow's API Cost. Caching Reversed the Result.
Jev cut one classification workflow's modeled API cost by 35%. Turning on prompt caching for the direct model reversed the result. Here is what to measure before adding a decision model.
I Replaced mem0 With a Self-Built Shared Memory System, and Measured It
Six months after calling Claude Code memory and mem0 complementary, I run neither. One human-reviewed vault now serves every AI tool I use, and adding local embeddings to keyword search moved the top-8 hit rate from 89% to 97%.
Can You Trust an AI Spending Limit? What My Gateway Test Showed
A small gateway test stopped repeated requests without cutting off another key. The allowance still overshot. Here’s how I’d turn that result into a buying decision.
Why Per-Tool Permissions Aren’t Enough for AI Agents
I tested prompt-only, stateless, and AgentCore temporal controls to show why AI agent permissions can still allow an unsafe sequence of tool calls.
How I Used AI to Reduce a Manual Matching Review Queue
How I added an AI operations review system to an existing matcher, supported 62 of 337 blind cases, and kept uncertain decisions with people.
Why I Moved MCP Server Authentication Off n8n—and Kept the Workflows
I moved MCP server authentication from n8n to AWS while keeping proven write workflows. Here is the architecture, security boundary, and migration tradeoff.
How a Quoted $7K/Year Supabase Access Path Became a $38–$45/Month AWS Read Layer
A quoted $7,000-per-year access path led me to build a governed AWS read layer estimated at $38–$45 per month—without replacing Supabase.
Codex vs Claude Code: The UI Is the Difference
Codex vs Claude Code is less about raw intelligence and more about workflow. Codex feels better when you are juggling real projects.
Microsoft Work IQ API: I Hit a Wall Nobody Warns You About
I configured the Microsoft Work IQ API end-to-end and still hit a 403. Here’s what Microsoft’s new agent API actually takes to use.
The Hidden Cost of Self-Hosting AI: Not the Server Bill
Self-hosting AI looks like a bargain at $48/month — until it breaks and you realize the bill was the easy part.
Running an AI workflow that costs too much?
The AI Workflow Review finds the one change that makes it cheaper and better, in one week at a fixed price. Start with a 30-minute call.