BLOG.

Every car brand structures its parts catalog navigation differently. Volvo goes four levels deep, Nissan has a two-phase BOM. Here's the parser architecture that handles all of it without turning into a mess.

CoreSynth now syncs AI model prices automatically from the upstream catalog. It doesn't just overwrite numbers — it guards against price spikes, writes an audit trail, and separates estimates from reality.

Why login history in the J+P Autodíly admin panel gets its own table instead of sharing the audit log — and why the login record must never touch the critical path.

Why having an e-shop isn't enough — and how a single generated XML feed puts a small bike shop's catalog in front of shopping comparators, twice a day, with no manual work.

Why we removed the Filter button from the J+P Autodíly admin panel and let filters write directly to the URL. Debounce, server-rendered listings and progressive enhancement in one shared component.

Why a delivery note per order wasn't enough — and how a single page grouped by delivery routes, one button per route, and an order timeline transformed warehouse operations.

Why 'just refresh the page' doesn't cut it when an admin monitors orders in real time — and how a single SSE pipe with a fingerprint replaced ten polling queries.

Why typing part numbers on a phone keyboard while standing at a pallet is the wrong workflow — and how a mix of native scanning, WebAssembly and server-side OCR turns a phone into a receiving terminal.

Why storing a few numbers in a database isn't enough — document state transitions, correct VAT splitting from gross prices, and a unified outstanding balance for both admin and customer.

An AI agent needs access to tools — but not all tools are equally risky. Orca introduces per-tool permissions with a YOLO mode, giving each tool its own trust level.

Complete redesign of the Sarah Hair Salon homepage — from a generic SaaS template to an editorial experience that actually looks like a salon.

How I added automatic language switching to my portfolio based on timezone and browser language — entirely client-side, no API, no flicker.

Mission planning in Orca no longer forces a sequential pipeline. The planner now builds a dependency graph where independent phases run in parallel and actually use the available agent capacity.

Orca now sends web push notifications to your phone. When an agent needs your input, a phase needs approval, or a mission finishes, you get a notification with action buttons right in it.

Every task in Orca can now have its own GitHub PR setting: default, on, or off. Two bugs were also fixed that allowed shipping a partial PR before the mission was complete.

Every stylist now gets her own ICS link for bookings. Connect it to Google Calendar, Apple Calendar, or Outlook and you'll see all your appointments without logging into the admin panel.

Alex generates massive system prompts — hundreds of thousands of tokens. OpenAI automatic prefix caching is free, but only when the prefix stays stable. Here's how a single timestamp bug was killing the cache, and how we fixed it.

AI tools are powerful, but each one usually lives in its own thread. So I built a shared memory layer that connects Claude Code, Codex and Hermes into one long-term context.

I stopped showing users raw numbers and gave them progress bars, a clean token breakdown, and a modal with cost projections instead. Here's how and why.