Inspiration
Growth systems usually start after a decision: a CRM records a lead, a dashboard counts activity, and a finance report arrives later. The risky step happens earlier—deciding which opportunity deserves time or money, what evidence is trustworthy, and when an autonomous system must stop.
Revenue OS makes that decision inspectable.
What it does
Revenue OS turns uncertain commercial ideas into small, falsifiable experiments. It keeps:
- dated evidence with source, publication date, retrieval date, confidence, and decision consequence;
- a scored opportunity pool with exactly three finalists;
- a risk card before each action, including maximum loss, privacy, legal, delivery, reputation, stop, and rollback conditions;
- predicted conservative, baseline, and optimistic outcomes beside verified actual results;
- an audited ledger that separates capital, confirmed revenue, receivables, spend, refunds, tax, and profit.
The no-login Build Week judge path includes an interactive decision gate. Change canonical status, competition, payment evidence, pre-revenue cost, privacy, and legal authority to see the same action become EXECUTE, OWNER GATE, or STOP.
The production system also includes a local-first Shopify CSV risk scanner. Files stay in the browser: the scanner identifies structural risks without uploading customer data. Fully synthetic samples show how ambiguous rows remain on HOLD instead of being guessed into a result.
How we built it
The project was created during the Build Week submission period with Codex and GPT-5.6. Codex and GPT-5.6 translated the operating objective into schemas, UI, Cloudflare D1 persistence, local-first analysis, risk controls, tests, deployment, and an iterative evidence loop.
The stack uses Next.js and React with TypeScript, Vite/vinext, Cloudflare Workers, Cloudflare D1, and a public sanitized GitHub repository. The live app is deployed with no judge credentials required.
Key decisions and challenges
The hardest problem was not generating more activity. It was preventing activity from being mistaken for value.
- Page views, deployments, form availability, quoted prices, and intent are never counted as revenue.
- Web content is treated as untrusted evidence, not instructions.
- Missing or conflicting evidence produces HOLD or STOP, not a confident guess.
- Funds, identity, contracts, tax, sensitive data, and irreversible actions remain explicit human gates.
- Failed channels are retired by written stop conditions instead of being retried under new wording.
Codex accelerated implementation, research, testing, and reversible operating work. The human owner set the capital limit and the final authority boundary. GPT-5.6 helped turn observed failures into durable operating controls rather than one-off patches.
What we are proud of
Revenue OS is a coherent working product rather than a static dashboard. The public repository contains setup instructions, synthetic sample data, tests, dated implementation history, and a clear account of how Codex and GPT-5.6 shaped the result. The production app exposes the decision gate, operating dashboard, and privacy-preserving scanner without login.
What we learned
Autonomy improves when stop conditions are designed before execution. A useful agent should not merely act faster; it should preserve evidence, distinguish cash from intent, expose uncertainty, and know which decisions it does not own.
What's next
Next steps are to evaluate the decision log against real commercial outcomes, measure prediction error, and keep improving only when qualified conversations, collected revenue, delivery margin, or risk reduction provide evidence. No deployment, test, or draft submission is treated as customer traction.
Built With
- cloudflare-d1
- cloudflare-workers
- codex
- gpt-5.6
- next.js
- react
- typescript
- vite
