Enterprises don’t buy demos. They buy reliability.

How I build

Six principles, each with a receipt

Everything I ship follows the same doctrine, formed by watching enterprise buyers say no to impressive demos and yes to boring reliability. Each principle names the pipeline stage where it does its work, and the case where you can check it.

1. 1 ·Start with the work, not the model.

   Before anyone names a model, I go and look at the job: who does it today, what it costs them, where it breaks. The copilot began as a consultant’s day-plus proposal cycle, not as a feature; the practice platform began with the observation that companies had bought AI licences and usage had plateaued at novelty. A product that starts from the work knows what good looks like before it generates anything.

   Seen in: Station 01 · The Problem · A practice platform, not a course

2. 2 ·Deterministic cores, LLMs at the edges.

   Anything that must not be wrong — costing, sequencing, data transforms, compliance — runs as deterministic code. The model handles what it’s reliably good at: language, synthesis, rationale. My journey-design engine generates nothing until seven deterministic steps and a human approval complete; the LLM then writes prose about decisions already made. I formalized this as a Workflows/Agents/Tools pattern: SOPs orchestrated by agents, deterministic tools for everything that must be exact.

   Seen in: Station 03 · The Build · From internal tool to platform flagship

3. 3 ·The human decides.

   Every system I ship has a hard human gate at the moment of judgment: the confirmation bar that surfaces every assumption before a proposal finalizes, the competency-approval gate before a journey exists, the metadata confirmation step before a client report generates. Autonomy is a dial, not a virtue — and in enterprise products, trust is the feature.

   Seen in: Station 04 · Trust · The deterministic gate

4. 4 ·Anti-hallucination is system design, not prompting.

   You don’t ask a model nicely to stop making things up. You build validators that hard-fail on uncited figures. You write adversarial test suites that attack your own confidentiality boundaries. You corroborate across four models and score consensus before believing a claim. Roughly 230 automated tests stand between my flagship product and its users.

   Seen in: Station 04 · Trust · From internal tool to platform flagship

5. 5 ·The moat is the knowledge, not the model.

   Models are rented; ontologies are owned. My platform work is grounded in a proprietary structure — 224 skills across 9 domains, 658 behavioural indicators, 19 journey templates — that makes generic models produce specific, defensible output. When the next model generation arrives, the moat transfers.

   Seen in: Station 02 · Discovery and the decision · From internal tool to platform flagship

6. 6 ·Know what not to build.

   Every roadmap I run carries an explicit don’t-build list with reasons. The strategy document I’m proudest of recommended against my own product’s standalone positioning. Sequencing beats ambition: internal users → partners → self-serve is how creation platforms actually win.

   Seen in: Station 05 · Launch and learn · Killing my own product’s positioning