Blog
Thoughts on AI, building products, and the occasional hot take.
1.6 Billion Tokens a Week on 120 Watts
Nine agents ran around the clock this week. Most produced garbage on Monday. By Friday, after four debug cycles, the architecture looked completely different.
Stop Hiring AI Agents. Start Hiring AI Employees.
We built an agent pool architecture with 4 workers and an ops agent running 24/7 on systemd. Swappable personality profiles, isolated state, pennies per day. Like a staffing agency for AI.
The SaaS Apocalypse
Software stocks are in freefall. The panic is indiscriminate, but the underlying disruption follows a clear pattern—and creates massive opportunity for builders.
First Light on the GX10 Cluster
The journey of deploying MiniMax M2.1 across two NVIDIA GX10 units. What broke, what worked, and what others can learn from it.
The Sandwich Defense
Prompt injection drops from 50%+ success rate to under 2% with one structural change. Here's the pattern and how to implement it.
Moltbot Gets a Brain
Building a local AI assistant to help run my business. The tech is ready and prices are going up.
Clawdbot With Teeth
Setting up Clawdbot on an RTX 5090 laptop instead of a Mac Mini, with plans to let it argue for its own hardware upgrade.
Effort Asymmetry
The competitive advantage isn't working harder. It's not needing willpower to do the work at all.
The Manhattan Project Needs Silver
AI is the new Manhattan Project. Every breakthrough needs physical infrastructure. Silver sits at the center of batteries, solar, military tech, and data centers—and we've run deficits for five straight years.
Operators Become Supervisors
Software devs stopped writing code and started directing agents. The same shift is coming to battlefield robotics—and it's coming there first, before commercial applications.
Let Go of the Code
One person can do what 50 did pre-AI. Stop touching code and let the agent do it.
Where the Work Goes
If 1 engineer can do what 50 did, what happens to everyone else? The bottleneck was never ideas—it was cost.
We Rebuilt Our Video Pipeline as an AI Agent
From brittle scripts to a self-correcting pipeline. How we turned video generation into a multi-phase agent that researches, writes, critiques, and improves its own work.
2.6TB Freed in One Evening: AI-Assisted System Maintenance
My workstation was 93% full with a forgotten 4TB NVMe sitting empty. Claude Code found it, set it up, and cleaned house while I watched.
Doc-First Development: Programming in Markdown
The architecture doc is the product. Code is just the implementation. Here's how we actually work with Claude Code.
Debugging AI Voice: Making It Stop Sounding Like AI
Our social media bot sounded like garbage. We couldn't explain why. Here's the agentic debugging process that's fixing it.
Adding Web Search to Your AI Agent
How we gave our AI real-time context with Brave Search and URL fetching. The integration details nobody talks about.
How We Cut LLM Costs 75% With a 2-Tier Architecture
The pattern: cheap models filter, expensive models generate. A simple architecture change that dramatically reduces API costs.
December: 176 Commits and Zero Lines of Code
What happens when building products becomes more fun than gaming. A month of shipping.
How We Build Software 10x Faster With AI
The workflow that lets us ship production software in days. The secret: stop typing code yourself.
Give Claude Code a Memory: Connecting RAG via MCP
How we built an MCP server so Claude Code can query our internal knowledge base instead of re-explaining context every session.
Introducing Nova Assistant
An AI assistant that runs in your infrastructure, keeps your data private, and costs a flat monthly fee.