I run five AI systems. They don’t share a platform. They don’t share context. They don’t talk to each other. I’m the routing layer.
This sounds inefficient. It’s the most efficient architecture I’ve ever built.
The Fleet
Each system does what it’s best at for the cost.
Claude handles deep work. Complex multi-file code, architecture, debugging, anything that needs long context and precise execution. Within Claude, I route between models — Opus for the hard problems, Sonnet for routine operations. Same infrastructure, different weight classes.
Gemini handles large documents and multimodal research. A million-token context window at their price point is unbeatable for ingesting entire codebases or long academic papers. I use it as a bulk research engine.
ChatGPT handles deep thinking and reflection. Not because it’s better or worse — because it’s different. Three models trained on different data with different objectives catch things one model alone misses. I proved this with a cold-drop probe: the same question sent to six AI instances with zero priming. The variance in what they caught was the signal.
Two specialized instances handle fleet coordination and overwatch. One monitors the other systems for drift. One watches the repositories for security issues.
Why Not One System
The obvious question. The answer has three parts.
Cognitive diversity. A single model has a single training distribution. It has consistent blind spots. When I send the same problem to three different architectures, the disagreements are more valuable than the agreements. Where they converge, I have high confidence. Where they diverge, I have a research question.
Cost architecture. Providers are building metering infrastructure to gate features and track usage. Single-provider dependence is a structural risk. When one provider tightens pricing, it affects one lane, not the whole operation. The cheap era of unlimited AI on flat subscriptions is ending. The response isn’t to find the cheapest provider — it’s to build an architecture where no single pricing decision can shut you down.
Failure isolation. When Claude has an outage, I still have Gemini and ChatGPT. When OpenAI changes their API, it affects one system. When any model has a bad day — and they all have bad days — the others hold the line.
The Routing Layer
The interesting part isn’t the systems. It’s the routing.
Every request goes through a decision: what does this task actually need? Deep reasoning? Bulk context? A different perspective? Routine execution? The answer determines which system handles it.
This isn’t automated. I’m the router. That’s deliberate. Automated routing optimizes for measurable metrics — speed, cost, token count. Human routing optimizes for the thing that matters: did this task get the right kind of attention?
When I’m debugging a subtle architectural issue, I want the model that will sit with ambiguity and trace implications three layers deep. When I’m processing a 200-page PDF, I want the model with the biggest context window. When I’m questioning my own assumptions, I want the model that was trained on different data than the one that helped me form those assumptions.
A router that doesn’t understand the task can’t make that call. I can.
What This Is
Enterprise architecture principles applied to a single user. Redundancy, failover, separation of concerns, cost optimization, vendor diversification. The same patterns I’ve designed for organizations with thousands of employees, scaled down to one person.
The difference is that at enterprise scale, you’re designing for the average user. At personal scale, you’re designing for the specific user. Every routing decision, every model selection, every prompt pattern is tuned to one person’s cognitive architecture. That’s not a limitation. That’s the feature.
The system compounds. Every session adds context. Every project refines the routing heuristics. Every mistake teaches the fleet something specific about how this particular human thinks and works. After months of daily use, the infrastructure knows things about my workflow that I couldn’t articulate if you asked me.
That compounding is why first-mover advantage in personal AI infrastructure isn’t measured in features. It’s measured in depth of self-knowledge encoded in the system.
For more on building infrastructure that thinks with you: subscribe on Substack.