PaperTrade

The brief described a real stock-trading app — real money, exchange routing, KYC/AML, broker-dealer licensing and paid market-data feeds. That is a regulated business, not a weekend build: six figures, months of compliance, and impossible to stand up on free tooling. The easy move is to quietly scope it down without saying so — or take the money and hit the wall later.

We did the honest thing first: separate what's buildable from what's regulated. A paper-trading simulator — virtual cash, simulated fills against real (delayed) quotes, live P&L — needs none of the regulated machinery and teaches the same product. That moved the real engineering to where it belongs: a money engine that stays correct under concurrency, can't be tampered with from the client, and survives an AI agent hammering it in a loop.

And it had to cost nothing to run. Free-tier quotas are small and unforgiving — a careless architecture burns through them in hours once bots are trading. The honest test was whether we could ship a real, secure, multi-platform product on $0 infrastructure without trading away correctness or governance to get there.

One Expo + React Native codebase targets iOS, Android and web, and we wrapped the same build as a native desktop app with Electron. The backend is Supabase — Postgres, Auth, Edge Functions and cron — with the trading engine living inside the database as SECURITY DEFINER functions, so every fill is atomic and server-authoritative.

Security is structural, not bolted on. The engine tables are read-only to users (row-level security plus explicit grant revokes); every mutation flows through an Edge Function running as the trusted service role, which calls the engine function. The ledger is append-only; orders are idempotent so a retried or duplicated request can never double-fill; daily limits are enforced atomically under a row lock; and a kill switch halts a runaway account at fill time, not just at submit. A bot API with per-account tokens and risk limits turns the whole thing into a safe sandbox for AI trading agents.

The architecture is what keeps it free: reads hit a price cache directly with zero function invocations, writes go through Edge Functions, and a single scheduled job owns every market-data call. We built it AI-accelerated with multi-agent workflows, then ran an adversarial review pass that caught real bugs — a missing service-role grant and two token-scope holes — which we fixed and re-verified before sign-off.

What exists is a complete, working product across four surfaces from one codebase: sign-up and auth, live-ish market data, candlestick charts, a full order ticket (market and limit), a portfolio with live unrealised P&L, order history, watchlists, multiple accounts, per-account risk limits and a kill switch — plus the desktop app. It's verified end to end: the app type-checks clean, all ten Edge Functions pass Deno checks, the 32-assertion pgTAP suite passes on a live Supabase stack, and an HTTP integration test proves the full path from auth through the Edge Function and database function to the ledger.

Commissioned cold, a secure full-stack trading product like this — a hardened money engine, a bot API and a desktop build — is a £90k–£140k agency engagement over three to five months. We designed and built ours AI-accelerated in a fraction of that, and it runs at $0 on free infrastructure.

That's the proof point. PaperTrade shows CIA ships governed, financial-grade product fast: correct under concurrency, tamper-resistant by design, safe for autonomous agents and cost-disciplined to zero. When we build for your business, this is the standard — and the speed — you can expect.