01 — Category
Notes from the workbench.
Build is where I keep the design, build, ship and improve loop for the services I make under Aulvem, plus the technical notes and coding basics that came up along the way. Honest about what got stuck, what was actually used, and why.
GA4 put Singapore at the top of my country report: 210 sessions in 28 days, one engaged, zero seconds of engagement. Cloudflare's free plan will not show the ASN, so I used what it does show — the screen size GA4 records and the Chrome version that fires the analytics beacon — and compared it with my readers in Japan.
I put two Cloudflare WAF rules in front of credential scanners impersonating AI crawlers. A month later the blocked requests were still in my analytics as 403s, the scan count had gone from 145 to 991, and the scanning had moved onto the names real crawlers use — ChatGPT-User and Amazonbot. What the numbers say and how I split the two jobs.
Anyone can send a request calling itself GPTBot. Splitting eight days of Cloudflare logs by verifiedBotCategory and by whether the path exists in my sitemap left 468 genuine article fetches against 991 credential-scanning requests. The same UA string splits across verified and unverified rows, and one of the agents in my logs cannot exist at all.
I was one command away from unpublishing five cross-posts to fix a suspected duplicate-content problem. Measuring the actual line overlap first killed the plan: 795 of 1,658 lines matched, but 762 of those were inside code fences and only 3 prose lines matched across 22 files. Here is the measurement and the canonical support matrix that replaced the deletion.
Export a Google Sheet to .xlsx and ARRAYFORMULA / FILTER / SORTN / QUERY survive the trip, even though Excel has none of them. Worse, the IFERROR that was already wrapping them swallows the missing-function error, so the screen shows 'no matches' instead of #NAME?. Here's how I counted the broken cells with openpyxl and rewrote them as helper columns plus INDEX/MATCH so they run on Excel 2016.
Sending a notification the instant you detect the condition means stock or price can change between detection and delivery, and the alert goes stale. Instead, detection only enqueues into an outbox; a dispatcher re-fetches and re-checks right before sending, and drops anything that no longer holds. Here's the idempotent enqueue, the send-time re-check, and exponential-backoff retries as a state machine on Cloudflare D1.
Cloudflare Workers' free plan caps cron triggers at five per account, which runs out fast once several services share it. Here's collapsing everything into a single per-minute cron that branches on UTC time to run patrol, send, and daily maintenance — and fitting each tick inside the 50-subrequest free-tier limit.
To send Web Push without accounts or email, you have to decide what holds the recipient before you send anything. Here is how an anonymous device token, minted into a cookie and delivered from Cloudflare Workers with VAPID over webcrypto, handles subscriptions, revocation, and the iOS and FCM gotchas that only surface on real devices.
Rakuten's item search API never returns the JAN barcode, while Yahoo! Shopping looks it up directly via jan_code. Building a cross-marketplace price watcher on that asymmetry, I split product identity into three certainty tiers — JAN match, name fuzzy match, no match — with the scoring weights, thresholds, and the notification rule that keeps mis-merged items from firing alerts.
Building a price watcher on my own, the calculation I underestimated most was 'display price minus the points you earn'. Splitting reward points into three certainty layers (confirmed, conditional, assumed), flooring each layer before summing, ranking on the confirmed layer only, and why shipping stays out — the design decisions I landed on, with code.
How I generate llms.txt and llms-full.txt as Astro API routes, driving them from Content Collections and splitting English and Japanese with two axes — filterLang (which posts to include) and docLang (the document's own language) — plus how it sits next to robots.txt.
Astro's i18n routing doesn't emit hreflang on its own, and a naive 'always emit all locales' implementation points at 404s for one-sided posts. Aulvem uses Content Collections helpers to emit alternates only for languages that actually have a published entry, then handles the reciprocal and x-default rules on top.