Traditional AI product
$0.01 - $0.50
per interaction
Margins shrink with every user. AI bill grows with traffic.
Back to home
Methodology
How we build products that do not need AI to run.
Four phases. AI does the heavy lifting while we build, so what you receive runs as plain software: the same answer every time, no live model call on the user path, no per-use AI bill.
The four phases
The workflow from raw idea to owned system.
- 01Spec before prompt
Strategy
Define the problem, the user, and the boundary of what we will and will not build. Write the rule we are going to ship, not the prompt that simulates one.
- 02Prototype judged by humans
Design
AI-assisted, human-judged. Stitch and Google AI Studio for high-fidelity prototypes; reviewed for hierarchy, accessibility, and brand fit before code starts.
- 03Agents execute explicit specs
Build
Claude Code and Codex write production code under explicit specs. Tests are part of definition-of-done, not TODO.
- 04Source and keys handed over
Deploy
Vercel for frontend, Railway for Python backends, Supabase for data and auth. Source code, brand assets, deploy keys, and operating notes are transferred.
Worked Example
The CRS Calculator, four hours, end-to-end.
IRCC publishes the Express Entry CRS scoring tables. They do not change. The calculator at /tools/crs-calculator is a faithful encoding of those tables in TypeScript.
- 0:00
Pulled the CRS tables straight from the IRCC PDF. Claude transcribed each scoring section into JSON; we cross-checked against three sample profiles.
- 0:45
Wrote the scoring functions: age, education, language (CLB), work experience, spouse adjustments. Each gets a unit test.
- 2:00
Built the React form. Sliders for age + CLB, dropdowns for education + experience, and a result chip that updates on every keystroke.
- 3:30
SEO meta + JSON-LD WebApplication schema. Sitemap entry. Plausible tracking on result computation.
- 4:00
Deployed. Runtime path: browser compute, no model call per calculator run. Even if a million people use it tomorrow, the bill stays the same.
The artifact, running here
The calculator that timeline built.
Every keystroke below computes in your browser. No model call, no server round-trip - which is the whole cost model, demonstrated.
The cost model
Deterministic vs. traditional, at scale.
Rules-first product
Near-zero
per interaction
AI builds the rules at design time. Runtime is pure compute.
Illustration at $0.05 per model call vs roughly $25 flat hosting. Arithmetic, not a quote.
Per-interaction AI product$2,500/moclimbs with every user
Rules-first product$25/moflat, at any traffic
AI billRules-first0 → 50,000 interactions
We use AI to build sophisticated rule engines, templates, and computation systems. Then they run forever at near-zero cost. You own everything. The slider above is itself a rules-first build: pure browser arithmetic, no model call.
Stack and tools
What runs where.
Every AI tool sits on the build side of the line. What ships is a conventional stack, boring on purpose. That is how the cost model holds.
Build timeAI, in our hands
Strategy + writing
Claude - ChatGPT - NotebookLM - Gemini Deep Research
Design
Stitch - Google AI Studio - Figma
Build
Claude Code - Codex - Antigravity (parallel agents)
Review + audit
Codex cold-eye audit - ChatGPT red team - GitHub Actions
RuntimeIn yours
Framework
Next.js 16 - React 19 - Tailwind
Data and auth
Supabase Postgres with row-level security
Hosting
Vercel for the front end - Railway for Python services
No model call on the user path. Where a feature is genuinely interpretive it is flagged, premium-only, and capped per user per day.