Senior full-stack developer — applied AI and web platforms
Professional context
One company since 2018, on client work rather than internal demos — and the work moved with the stack. Web and platform engineering first, applied AI now: assistants that sit inside real workflows, recommendation systems, and the evaluation that tells you whether either actually works.
The work
What the work actually is
I've been at IWConnect since March 2018, always on client-facing systems rather than internal demos. The role moved with the stack: it started in web and platform engineering, and it now sits mostly in applied AI — assistants, retrieval, recommendations, and the evaluation around them. Full-stack here means I own the whole thing: the API and data model underneath, and the interface people actually touch.
Applied AI
The assistants I build do one job inside a workflow that already exists, rather than answering anything about anything. An email helpdesk responder triages incoming support mail and drafts replies. An insurance policy renewal assistant handles the renewal path. In both cases the hard part wasn't generation — it was routing, grounding the answer in the right source, and deciding when the system should stay quiet instead of guessing.
Recommendations, and measuring them
I built a product recommendation engine, and then an SDK for measuring recommendation quality in a repeatable way. The second one is the one I'd defend: a recommender you can't measure is a recommender you can't improve, and "it looks reasonable" is not a metric.
Email and marketing infrastructure
A long stretch of email engineering — MJML templates, SMTP delivery, a Litmus-based sandbox workflow for validating rendering before send, and a preview tool for checking HTML email across clients. Unglamorous work with brutal compatibility constraints, and where I learned to treat rendering and QA as part of the build rather than a step after it.
WordPress platform engineering
Theme and plugin development, custom post types, and Elementor builds for client sites, plus a generator that turns a design-token system into WordPress-compatible HTML. Years of watching hardcoded post models age badly is exactly why Alex CMS has no hardcoded content models — every type is data, defined at runtime.
What I took from it
Working across all of those layers is why the systems I build now are whole systems rather than features. The retrieval layer, the content model, the cost tracing and the interface are all mine, and they're built to be handed over without me.
- What I did there
- Email helpdesk responder — triages incoming support mail and drafts replies
- Insurance policy renewal assistant — AI-assisted handling of the renewal workflow
- Product recommendation engine, plus an SDK for measuring recommendation quality repeatably
- Email infrastructure end to end — MJML templates, SMTP delivery, Litmus sandbox validation, cross-client preview tooling
- WordPress platform engineering — themes, plugins, custom post types, Elementor builds for client sites
- Design-token → WordPress HTML generator, the first version of what became Theme DNA