Build Log
What was planned vs what was built.
From a weekend project to a governed intelligence infrastructure in seven months.
A small model running on a gaming rig that could help with Excel logic and answer questions without relying on a cloud service.
"Let's build a baby Dex who can speak spreadsheet and still be charming."
Build weekend scheduled: August 10–11, 2025
By March 2026, the system looked nothing like the August plan. It looked like this.
Each divergence tells a story about how systems evolve when you follow the problem instead of the plan.
GPT-OSS never materialized as a practical local option. The operator discovered Ollama and its ecosystem of quantized models. A 7B model on 8GB VRAM runs fast enough for real work. The 20B plan assumed bigger was better. The build proved that governance matters more than parameter count.
LM Studio is a GUI. Ollama is a server. The moment the operator wanted to run inference from another device — a phone, a laptop, a second machine — the project needed an API, not an interface. Ollama exposes a REST endpoint by default. That one capability unlocked remote access, multi-node inference, and AutoCouncil orchestration. None of which were in the original plan.
The original goal was bidirectional Excel read/write. That happened — but it became a small piece of a much larger system. The operator's archive turned out to be far more valuable than any single spreadsheet. ChromaDB made it searchable. The RAG pipeline made it useful. Excel integration still exists but it's not the center of the system anymore. The corpus is.
The August plan had a one-paragraph personality prompt: "Be helpful, clear, a little witty." By March, the system prompt had been rewritten four times, tested against three documented failure modes, reviewed by nine AI models, and hardened with 18 explicit anti-pattern rules. The personality layer became a governance layer. "Charming" was replaced by "governed."
The operator discovered that the output of the system was publishable. Not as a product demo — as actual content. Council deliberations, knowledge graph exploration, constraint documentation, operator reflections. The "GUI" became a website. The website became a publication platform. The publication became four distinct series.
The original vision was a single model that could be encouraging. What emerged was a multi-model system where ten different AI platforms independently review every major architectural decision. The council doesn't compliment. It challenges. It produces LOCK, REVISE, and REJECT verdicts. It finds what's missing. The operator doesn't need encouragement anymore. He needs governance.
The system couldn't be built in a weekend because the operator didn't know what the system was yet. Each component emerged from the one before it. The RAG pipeline created the need for governance. Governance created the need for the council. The council created the need for a publication process. The publication process created the need for series architecture. None of this was predictable from the August plan. All of it was inevitable once the building started.
Every divergence follows the same structure:
This is Chaos → Structured → Automated applied to the development process itself. The plan was the chaos. The build was the structuring. The automation is still emerging.
One thing. The most important thing.
The plan said: build something local that helps you think.
Everything else changed. That didn't.
Dex Jr. helps the operator think. The corpus holds the memory. The council challenges the reasoning. The publication makes the thinking visible.
The baby Dex who speaks spreadsheet became a governed intelligence infrastructure.
But the job is the same.