CIA Client Portal

Consultancies bleed time into email threads: where's the latest report, what's the project status, who's handling the invoice query. A portal fixes that — but a portal is also exactly where software risk concentrates, because sign-in, sessions and file upload/download are the classic attack surface. It's the last thing you want built carelessly.

The first draft of this one was built by an autonomous AI agent team working a board ticket. It looked complete — pages, tables, buttons, the lot. "Looks complete" is precisely the trap with AI-built software: underneath were hardcoded tables, placeholder alerts and pages that never spoke to a real API. The interesting question isn't whether AI can draft a portal in an afternoon — it can — it's what it takes to make that draft genuinely trustworthy.

And it had to graduate from somebody's terminal to something client-facing in behaviour: surviving reboots, restarting itself if it dies, with logs and a health probe — a service, not a demo that lives as long as the laptop lid stays open.

We audited before we polished: four AI agents reviewed the draft independently and adversarially, and between them surfaced 55 findings across security, correctness, accessibility and operations. That list — not the feature list — set the work order, and the server was near-rewritten around it.

The hardening is specific and checkable. Sessions use crypto-random identifiers with an enforced 8-hour expiry, periodic eviction and genuine server-side logout invalidation; sign-in is rate-limited per IP; uploads go through real storage with sanitised filenames and size caps, and downloads and sharing require an authenticated session. Input lengths are bounded, the error handler never leaks a stack trace, and even the 404 is branded. On the front end every page was wired to its real API — no stub tables left — with all interpolation routed through an escape guard against XSS, the session re-validated on every page load, and accessibility brought up to AA: landmarks, skip links, contrast and 44px touch targets.

Then we proved it and made it permanent: 200 concurrent requests without error, an 18-for-18 smoke suite, the full browser journey — sign-in to dashboard to tickets, files and resources — with zero dead links. The portal now runs as a supervised macOS service with keep-alive (kill it and it respawns, which we verified), logs on disk and a health endpoint for liveness checks — restyled throughout to the CIA brand with no CDN dependencies.

What exists is a working portal doing the actual job: sign in, see projects with progress and activity, upload and share real files, raise and track tickets, browse the shared resource library — running for days as a supervised service rather than a process in someone's terminal.

We'll be honest about its scope: it is deliberately a demonstration build. Credentials are demo accounts and the data store is in-memory by design; wiring in a real user store, a database and a content-security policy is the documented next step, not a hidden gap. The hardening pattern — audit, rewrite, verify, supervise — is the part that carries to production unchanged.

The meta-story is the point. Our delivery model is AI does the drafting and governance makes it shippable — and this build shows the second half working: 55 findings caught by our own adversarial audit before a client ever saw the portal. That's the gap between "an AI built it" and "you can trust it" — and closing that gap is exactly what we sell.