Approach
How the work gets done.
Five stages, the same every time. AI is involved at each one, and so is a person. The split is the whole method: the model does the typing, the founder keeps the judgment.
- 01
Find the gap
Industries software skipped.
Dance studios still run on paper waivers. Group travel still runs on screenshots. That is the opening. Each project starts by shadowing how the work is actually done today.
- Perplexity
- Claude
- Conversations
- 02
Model it
Schema before screens.
The data model and the permission rules get written first, in SQL, with row-level security on every table. The screens are a view on the model, which keeps the product honest when features pile up.
- Supabase
- Postgres
- SQL
- 03
Generate, then rewrite
AI builds the first version.
Claude Code produces a working version of each feature in hours, one pull request at a time. The parts that have to last, auth, permissions, payments, real-time sync, are rewritten and reviewed by hand. The rest ships as generated.
- Claude Code
- Cursor
- Lovable
- 04
Harden
Types, analysis, review.
A typed client generated from the schema, static analysis on every pull request, and a real-time hook per entity so UI state never drifts from the database. Eval suites cover chat, document parsing, and the MCP server.
- TypeScript
- SonarCloud
- GitHub
- 05
Operate
Run it as a business.
Ship as a beta, release weekly, keep API keys in edge functions and infrastructure cost near zero. Then listen to the people using it and go back to step one.
- Replit
- Supabase
- PostHog
Principles
Own the whole thing
Product, code, data, support. No hand-offs means no gaps where a decision has no owner.
Permissions are the product
A studio that sees another studio's roster is a failure, whatever the UI looks like. Security rules are written first and reviewed like code.
AI does the typing. A person does the thinking.
Models are fast at the first version and bad at knowing which parts matter. Judgment stays with a person.
Cheap to run
Serverless by default. Costs scale with users.
Tooling
What the firm builds with, and why.
The list changes. The rule does not: a tool earns a place by shortening the path from brief to beta without making the product harder to run.
- Claude
- Agentic coding and research
- Claude Code
- Builds each feature, one pull request at a time
- Cursor
- AI pair-programming in the editor
- OpenAI
- Design specs for WanderLuxe's Print Studio
- Gemini
- In-product chat and document parsing
- Supabase
- Postgres, auth, realtime, edge functions
- Replit
- Hosting and rapid prototypes
- GitHub
- Source control and review
- SonarCloud
- Static analysis on every pull request
- Lovable
- First screens, fast
- Perplexity
- Research
- PostHog
- Product analytics
- Stripe
- Payments and subscriptions
- Twilio
- SMS sign-in codes
Working with Reminiscent
The same method, applied to someone else's problem.
Reminiscent shows companies what AI can do inside their business, builds the first version with the people who will use it, and hands it to their technical team to roll out. Nobody gets buried in new tools along the way.
Fix a specific problem
One workflow that eats your team's week, scoped tight and rebuilt with AI.
Find where AI helps
A pass through how the business runs today, ending with the short list of workflows where AI saves the most hours.
Build the first version
A working version built with the people who will use it, pairing their know-how with new AI tools. Your technical team scales it from there.