A published testimonial is worth exactly what its verification is worth, and most
sites verify nothing. The interesting problem was not the sign-in flow, it was deciding
where verification stops.
Trust and identity · Production · 3 min read
Most websites expect an agent to scrape them. This one registers tools a capable
browser can invoke directly, and the build fails if the published list and the
registered tools disagree.
Agents and machine access · Production, on a time-limited origin trial · 3 min read
Two walk-throughs of this site's own machinery: what happens when you press send on
the contact form, and what an AI agent sees when it arrives. The failure paths are
shown rather than implied.
Reliability and cost · Production · 4 min read
Adding a second sign-in provider is where verification systems usually go wrong. The
fix was structural rather than careful.
Trust and identity · Production for two providers, implemented for a third · 4 min read
Bursty work on a host that stayed awake regardless. The obvious move was also the
last step rather than the first.
Reliability and cost · Production · 4 min read
A scheduled registry purge deleted the image production was running on. The obvious fix had a hole in it, and the hole is the interesting part.
Reliability and cost · Production · 3 min read
Mail providers send daily reports on who is sending email as your domain, and almost nobody reads them. The ingestion was the easy half; knowing what your DNS actually publishes was the half that mattered.
Security and secrets · Production · 4 min read
Every sending platform keeps its own list, and the obligation does not migrate when you change vendor. Most of the work was database permissions, not cryptography.
Trust and identity · Production · 4 min read
Eight of twenty-six jobs failed and the VM was deleted after 32 hours. What the failures actually said, and what changed when it came back four months later.
Reliability and cost · Production · 3 min read
A proxy with 170 routes, fifteen pointing at endpoints that were never there. Nothing caught them, because a compiled integration is not a verified one.
Integrating other people's systems · Production · 3 min read
There is no undo on a shared mailbox. Preview before apply, a safety cap, and going live only on an explicit go-ahead.
Integrating other people's systems · Production · 3 min read
A dynamic IP against a locked-down key vault. The real alternative was never “keep running the script”, it was “widen the rule and forget”.
Security and secrets · Production · 3 min read
Four services moved off long-lived keys onto managed identity. The code change is trivial; the failure mode moving from startup to first use is not.
Security and secrets · Production · 3 min read
Three releases where the pipeline was wrong about production. Deployment success, service health and build identity are three different claims, and only one of them notices the wrong build.
Reliability and cost · Production · 5 min read
A heading read as an invoice number, a repair that swapped one false value for another, and a nightly sync that turned errors into empty lists and called it success.
Integrating other people's systems · Production · 5 min read
A scheduled email that never runs throws no error. Five times a twice-daily digest stopped, lost a section or nearly sent twice, and nothing raised an alarm.
Reliability and cost · Production · 6 min read
Writing the screens was the easy half. Five failures that only appeared after a green build, on a device, at the store or at sign-in, each with an error pointing at the wrong layer.
Integrating other people's systems · Production · 6 min read
A daily sync was refused every read for six weeks and logged success each time. Found because the fix for a different bug was not allowed to close until a real record showed it working.
Integrating other people's systems · Production · 5 min read
Five ways my own dashboard misled me without a single query failing, and a contact form that seemed to reject almost every submission, which was not the problem it looked like.
Reliability and cost · Production · 5 min read
A .NET security advisory broke my builds overnight with no code change. Why the right fix differed in each of four repositories, and why audit stayed on.
Security and secrets · Production · 5 min read
My hosting config was 98 bytes from its cap. 79 of its rules restated the platform, and the redirects among them were dropping ad click IDs.
Reliability and cost · Production · 6 min read
Every custom GA4 event on my site went silently undelivered after a move to the GTM Google tag. How one probe found it, and the lint that stops it.
Integrating other people's systems · Production · 5 min read
A backend port silently dropped one UPDATE and a column went empty for seven weeks. Why the backfill rule was tested on known answers before it ran.
Reliability and cost · Production · 6 min read
Case studies
Consent-aware first-party analytics with server-side channel classification and enrichment.
Trust and identity · Case study · 2 min read
Three fully separate site variants — business, recruiter, freelance — independently indexable and routed by preference.
Reliability and cost · Case study · 2 min read
Live Outlook availability as a queryable API — built so booking systems read the real diary, not a copy.
Integrating other people’s systems · Case study · 2 min read
Built to resolve duplicate enquiries into one canonical contact — matched, enriched, and synced to Microsoft 365.
Integrating other people’s systems · Case study · 2 min read
Server-side attribution built to survive ad blockers and report real conversions back to Google and Microsoft Ads.
Integrating other people’s systems · Case study · 2 min read
The site backend rewritten from Node.js to .NET 10 — a shared services library, its behaviour pinned by tests during the port, shipped through the same pipeline.
Reliability and cost · Case study · 2 min read
Built to take invoices from mail, OneDrive, and FreeAgent — extract, match, and write them back with full audit.
Integrating other people’s systems · Case study · 2 min read
GA4, YouTube, Google Ads, and Bing data ingested into first-party SQL behind a dashboard API.
Integrating other people’s systems · Case study · 2 min read
One Entra-gated dashboard — Windows, Android, and browser — for analytics, ads, leads, and security posture.
Integrating other people’s systems · Case study · 2 min read
Seven-phase billing automation with audit-logged writes, FreeAgent sync, and a WPF desktop app.
Integrating other people’s systems · Case study · 2 min read
A branded digest email twice each weekday: inbox counts, lead signals, site analytics and email security posture.
Integrating other people’s systems · Case study · 3 min read
Claude wired into DevOps, Microsoft 365, Azure, and ads platforms — with dry-run/confirm on every write.
Agents and machine access · Case study · 2 min read
DMARC and TLS-RPT reports ingested, archived, analysed, and alerted — automatically.
Security and secrets · Case study · 3 min read
The build-once pipeline, quality gates, scheduled operations, and post-deploy checks behind this site.
Reliability and cost · Case study · 3 min read
