At-home clinical testing

The client is a healthcare company running chronic care management programs, and it sells clinical testing that people take at home. Frank Robles has been its principal AI services architect since January 2026, and this platform is his by commit history.

At-home testing breaks in the gaps. The purchase sits in one system, the assessment in another, the clinician's review in an inbox, and the result goes out by hand. Every handoff is a place where a patient stalls, a clinician waits, or a minor's result lands with the wrong adult.

The business needed one record that runs from checkout to a licensed provider's review, with the consent rules enforced by the software rather than remembered by staff.

• Six condition-specific assessments. Cognitive and memory health, adult attention (ADD/ADHD), youth attention, sleep apnea (an OSA screener), concussion, and post-viral (Long COVID). A broad-spectrum assessment sits beside them.
• A validated test battery, launched automatically. Cognitive, attention, concussion and Long COVID assessments run on a validated cognitive test battery from a clinical assessment vendor, integrated by remote link.
• Intake, payment and results in one place. Stripe checkout, intake and result delivery live in the same application.
• Review by a licensed provider in the patient's state, on every session.

The patient picks the assessment, completes intake and pays by card. When the session is complete, the platform launches the right configured test battery for that screening, so nobody on staff has to send a link or pick a configuration.

Where a video step is part of the program, the patient is assigned their video and reminded. If they do not finish, the system retries on callback. Scheduled job services handle the reminders, the callback retries and the finalization of stale sessions, so an abandoned test does not sit open forever. A daily sync with the SMS provider keeps the messaging side current.

The result comes back to the patient. A copy goes to the patient's own doctor only where consent was captured. A minor's result routes to the parent or the purchaser, not to the child.

Review assignments go to clinicians by invite email. A licensed provider opens the session and reviews it.

The reviewer does not start from a blank page. The platform generates a draft clinical review with an AI model. If the model is unavailable, a deterministic fallback path produces the draft instead. The clinician works from the draft, and the queue never stops because a model provider had a bad afternoon.

Care navigators get depression and anxiety screeners with the escalation rule built in: PHQ-2 into PHQ-9 and GAD-2 into GAD-7. The short screen escalates to the full instrument by rule, so nobody has to remember when to switch.

A signed API serves the care portal, so the clinical side of the business reads results without anyone re-keying them from one screen into another.

This is a clinical product that is live and selling, so a bad deploy is not a cosmetic problem. The platform is containerized with development, staging and production profiles. On startup, development and staging run the migrations, a migration drift check, deploy checks, database checks, the seed commands and the full test suite. If a test fails, the build refuses to launch.

• Versioned releases. Semantic versioning, a written release checklist, and a changelog kept for every release.
• 1.0.0, production-ready, 26 March 2026. 1.2.0 followed on 6 April, 1.2.5 on 21 May, and 1.3.1 on 9 July 2026.
• Browser tests, not just unit tests. Playwright drives the real screens, and a visual regression checklist covers what automated tests do not.

Django and PostgreSQL. Mailgun for email with click tracking off, SMS, Cloudflare tunnels and scheduled job containers. Preview environments sit behind environment-controlled authentication with a branded access screen. Idle sessions log out.

• Django
• Python
• PostgreSQL
• Docker
• Stripe
• Mailgun
• Playwright
• Cloudflare tunnels
• SMS

Operations staff work in a formatted operations admin that sits in front of the Django admin. Day-to-day work happens in screens built for it; the raw admin stays for the few who need it.

Frank built it. The principal architect role here includes a second job: deciding which steps a person does and which a machine does, such as an AI review draft with a licensed provider's review on top of it.