The problems we get called in for, and how we work on them.

The situation. A few hundred open roles live in an ATS. The careers website is a separate CMS, updated by hand. Roles go live late, closed roles keep taking applications, and the same vacancy exists twice with different salary bands. Recruiters have started sending candidates the ATS link directly, which loses the employer brand and every analytics event with it.

The approach. Treat the ATS as the source of truth and stop writing job data anywhere else. Pull roles through the documented API on a schedule, validate each record at the boundary so a malformed or half-published role fails loudly instead of appearing wrong, and keep the site's own layer for the things the ATS has no field for: office pages, benefits, hiring-process copy.

What changes. Publishing a role becomes an ATS action, not a website ticket. Application tracking stays intact because candidates never leave the domain.

The situation. Two systems that were never meant to talk — say an ERP and a CRM — are reconciled every week by someone exporting CSVs and pasting them into a sheet with rules only they understand. It works until that person is on holiday. Nobody can say which system is right when the two disagree.

The approach. Map what actually moves today, including the manual steps and the undocumented exceptions, and confirm that reading with the person doing the job before writing any code. Then a middleware layer with the boring properties: an explicit field-level contract, idempotent writes so a retry cannot duplicate a record, a dead-letter queue for what fails, and logs an operations person can read without us.

What changes. The reconciliation stops being tribal knowledge. When a vendor changes their API, it breaks in a known place with an alert, not silently in next quarter's reporting.

The situation. A CMS and shop stack chosen when the catalogue was a tenth of its current size. Category pages are slow, every content change needs a developer and a release window, and the plugin that handles pricing has not been updated in years. A full rewrite has been proposed twice and cancelled twice, because nobody can afford the URLs and the checkout going dark for a weekend.

The approach. No big-bang cutover. Inventory what exists first — URLs, redirects, structured data, integrations nobody remembers — then move it in slices behind the same domain, keeping the old and new stack serving side by side until the last slice is verified. Redirect maps and search-visibility checks are part of the build, not a post-launch surprise.

What changes. Editors publish without a deploy. The migration becomes a sequence of reversible steps instead of one weekend everybody dreads.