For eleven years I worked on other companies' infrastructure. In January 2026 I left to build my own products for online stores.

I use AI daily, but I use it the right way. I have some simple rules to avoid slop and design that looks "ai-generated" (you know what I mean):

• 🧩 The model matters. Claude for code, Nano Banana for assets, DeepSeek for writing. That last one is a pure quality judgement, obviously, and has nothing whatsoever to do with DeepSeek being the one that still doesn't watermark its output. 🤫
• 🎯 Never let it get creative. A model is built to reproduce patterns, not to think, so originality is not a thing you can request. Work out the best approach yourself, then hand it over as the pattern to imitate.
• 🛡️ Guardrail everything. Every byte a model reads is untrusted input. Assume prompt injection is coming and make sure that when it lands, there is nothing behind it worth reaching.
• 🔍 Own the architecture. Read every line before it ships and understand why it is there. The model types faster than I do, but it is not the one accountable for the system.

AI chat that answers from your actual store.

Most storefront chatbots will cheerfully invent a price or a return window. ChatBuster syncs your live catalog, pages and policies, so every answer comes from real store data. It also handles "where's my order?" without anyone opening a support ticket.

• 🧠 Trained on your catalog automatically, with no uploads, no prompt files, and no API keys to babysit
• 📦 Real order tracking for guests and logged-in customers
• ✍️ One-line corrections that fix a wrong answer straight from the transcript, no retraining
• 🎯 Proactive campaigns on exit intent, dwell time or scroll depth
• 🌍 Replies in the shopper's language, automatically
• ⚡ Live in under two minutes on any theme

⭐⭐⭐⭐⭐ on the Shopify App Store and WordPress.org.

Ecommerce email marketing without the enterprise price tag.

Every incumbent charges you by the size of your list, including the people who already unsubscribed. AguMail bills only for contacts who actually want your email, and sends are unlimited on every paid plan. It runs on SES, where a thousand emails cost about ten cents.

• ♾️ Unlimited sends on every paid plan, priced on active contacts only
• 🛒 Ecommerce flows for abandoned cart, browse abandon, win-back, replenishment and cross-sell
• 🔮 Predictive segmentation using RFM, lifetime value and churn risk
• 💌 Deliverability suite with MX validation, disposable-domain screening and reputation monitoring
• 📊 Revenue attribution down to the flow and the product
• 🔌 Shopify & WooCommerce sync out of the box

Currently deep in deliverability and import tooling.

An SQS event bus for Medusa.

Medusa ships with an in-memory event bus, which means events die with the process and cannot be shared between instances. This module puts AWS SQS underneath instead, so you get at-least-once delivery, event grouping, and workers that scale horizontally. Community plugin, MIT licensed, published on npm.

I'm a polyglot programmer. I pick the language that fits the job, that the platform already speaks, and that has the most support behind it when something breaks at 2am. Over eleven years that has meant Java at Amazon, C++ at Google and Splunk, and TypeScript across everything I ship today.