Ayach Youssef

Case study — own product, live, 2026

ci-pulse

A hosted CI status board for small teams: sign in with GitHub, pick the repositories you care about, see every one’s latest Actions run on one page, and get a Slack or Discord message the moment a build turns red. Multi-tenant from the first commit, with a free and a paid tier.

Role

Everything: design, code, tests, security pass

Timeline

September — October 2026

Stack

Next.js · Postgres (Neon, Drizzle) · Auth.js · Stripe

Status

Live at ci-pulse.vercel.app. Source private.

Why this one

Start from code that already works

In September I wrote ci-status, a small Go CLI that prints the latest GitHub Actions run for a list of repositories and can wait for one to finish. It solved my own problem and nothing else. When I decided to build a product rather than take freelance hours, the fastest honest path was to turn that polling logic into a hosted service for teams, instead of starting from a blank idea.

The product is deliberately narrow. It does one thing a team of two to fifteen people actually wants: one page with every repo’s build status, and one message when a repo goes from passing to failing. Not a message per failed run, not a dashboard of charts.

ci-pulse — landing page
The ci-pulse landing page: the headline 'Know the moment a build goes red', a sign-in-with-GitHub button, and a mock status board showing acme/api passing, acme/web failing and acme/infra running, with a Slack alert underneath
The pitch and the product on one screen. The status board is a static mock here; the real one is below.

Architecture

Five decisions that shaped it

One OAuth flow, two jobs

Auth.js with GitHub as the only provider. The same sign-in that creates the session also grants the repo scope the poller needs, so there is no second “paste a token” step. GitHub offers nothing narrower that still reads private repositories, and the landing page says so.

Tokens encrypted at rest

Each user’s GitHub token is stored AES-256-GCM encrypted, packed as iv:tag:ciphertext in one opaque column. The key comes from the environment and is checked for length at boot. A database leak alone does not leak anyone’s GitHub access.

Teams, not users, own things

Eight tables: users, orgs, memberships, GitHub connections, repos, run status, webhooks, invites. Every repo, webhook and invite belongs to an org, and every server action re-checks membership and role for that org before touching a row.

The poller is not on Vercel Cron

Vercel’s free tier runs cron once a day; the product needs every five to fifteen minutes. So polling is a secret-guarded endpoint that any external scheduler can call, with a GitHub Actions workflow as the default caller.The Ship Log entry has the reasoning.

Cheap requests, polite backoff

Ported from the CLI and extended: each check sendsIf-None-Match with the last ETag, so an unchanged repo costs a 304 that does not count against the rate limit. A 403 or 429 raises a typed error carrying the reset time, and the batch stops instead of hammering.

Alerts only on the transition

The poller stores the previous status and fires a webhook only when a repo goes from passing to failing. Webhook URLs are validated against the real Slack and Discord hosts, so the server will not POST to an internal address someone typed into a form.

In the interface

Small on purpose

ci-pulse — dashboard
The ci-pulse dashboard for a team: one monitored repository, YoussefAyach01/YoussefAyach01.github.io, marked 'Waiting for first check', with a 'Connect repos' button and the note '1/2 repos on the Free plan'
A team's board, before the first poll. Plan usage is enforced server-side, not just shown.
ci-pulse — settings
The ci-pulse settings page: a Plan card reading 'Free: up to 2 repos, checked every 15 minutes, no alerts', an Alerts card explaining Slack and Discord alerts come with Pro, and a Team card listing one owner
Settings degrade honestly: when billing is not configured, the page says so instead of showing a broken button.

The security pass

What a checklist found on a “finished” app

Three weeks after the build was feature-complete I ran a full security checklist against it: input validation, authorization on every mutation, secrets, headers, webhooks, dependencies. Most of the earlier hardening held: no route lets one team read or change another’s rows, the cron secret is compared in constant time and fails closed when unset, webhook signatures are verified.

Three things did not hold, and each got a commit. The Next.js version in use had a critical remote-code-execution advisory in its Open Graph image helper and a server-side request forgery in image optimisation; both were patched by an upgrade. Environment variables were read ad hoc, so a missing or malformed value failed in the middle of a request; they are now validated once at boot by a single schema, and a bad deploy fails before it serves anything. And the original cron check used a template string, which would have accepted the literal header Bearer undefined when the secret was unset; that is the bug the constant-time check replaced.The full write-up is in the Ship Log.

Honest limits

What this is not

It is live and it has no customers yet. The code is typed, linted, audited and covered by 26 tests, and the sign-in, repo picker, settings and invite flows have all been driven end to end against a real database. The billing code is written, but I have not taken a real payment through it, and I would not claim otherwise.

The scheduler is the weak point. The polling workflow asks GitHub for a run every five minutes; on a quiet repository GitHub actually delivers one every few hours. The product’s promise depends on that interval, so the caller has to move to a scheduler that keeps time before anyone pays for the five-minute plan.The Ship Log entry has the measurements.

The repository is private while the product decision is open. I am happy to walk through the code on a call.