mkurri lab
AI experiments, one at a time
I Replaced a 702-Line ffmpeg Prompt With Remotion for AI-Generated Shorts. Here's the Diff.
My agents built shorts with ffmpeg drawtext at 576p and 134 kbps. Remotion plus one spec.json now renders 1080p reels and 16:9 videos. Numbers, spec, costs.
One VPS, 28 Vhosts, Zero Cloudflare Bill: The Stack Behind My AI Content Autopilot
28 nginx vhosts, 17 PM2 apps and 16 scheduled Claude jobs on one 8-vCPU VPS. The scheduler, the Cloudflare self-block fix, R2 pitfalls and IndexNow vs GSC.
AI Tool Articles Stay Wrong for 46 Days Before Anyone Notices
11 real staleness events from a Czech AI content site: median 46 days between a vendor change and my refresh pipeline catching it. Rot types, cadence, fixes.
Silent Success Is the Worst Bug: 8 Tools That Lied to My Agents
Exit code 0, HTTP 200, 'Uploaded.' Eight real cases where my agents got a success signal for work that never happened, and the guard I added for each one.
Approve Buttons Are Rubber Stamps: 8 Weeks of an AI Content Autopilot on Quality Gates
Approve buttons on a scheduled AI content pipeline turned into rubber stamps. Eight weeks of self-enforced quality gates instead: what they caught and missed.
My Cannibalization Detector Found 0 Duplicates. I Merged 24.
A title-similarity cannibalization detector flagged 0 pairs in 15 runs on my automated content site. Manual triage merged 24 pages. Here is why it failed.
Model Routing: How to Save on AI Without Losing Performance
Why automatic switching between models is the new standard for anyone building AI applications.
AI Security in 2026: Why Self-Hosted Sandboxes Are a Necessity
An explanation of why companies are transitioning to self-hosted solutions and how to protect corporate data when using AI agents.
How to Measure AI Efficiency in a Team? Tools for Observability
A practical overview of tools for monitoring the costs and performance of AI agents in your workflow.
Your CLAUDE.md Is a Liability, Not a Feature
The big global CLAUDE.md everyone proudly shares burns thousands of tokens before you type anything. Most of it is dead weight. Here is the audit, the cost, and the lazy-load pattern that fixes it.
Token Savings Guide
Why you're burning tokens faster than you think — and practical ways to cut your AI bill by 60%+ across any LLM tool or API.
Claude Code without Anthropic
Testing alternative AI providers behind Claude Code. What works, what breaks, exact configuration for Z.AI (GLM-5.1), Kimi (K2.5) and OpenRouter (Qwen 3.6).
Automated Content Pipeline
An end-to-end system where an AI agent orchestrates the full content lifecycle — from topic discovery to search indexing — across multiple publishing channels.
Claude Home Lab
Running Claude Code as 24/7 AI infrastructure. Always-on AI cofounder via Slack.
AI Image Model Benchmark
Comparing top AI image generation models. Same prompts, different results.
The End of the 'AI-enabled' Era: Why Agents Are the New Gold Rush
An explanation of the shift from conventional applications to autonomous agents that will manage your business for you.