At work
Deep in wiring the Mevrik AX Platform frontend to its live backend — the operator console, its omni-channel inbox and CRM, the AI agent studio, and the embeddable chat widget, all on the shared UI kit. The platform runs customer experience with AI agents while humans supervise; as frontend lead my job is making that fast and honest in the browser: real data over mock, verified in the page, never faked green.
At night
Two systems, feeding each other. The second brain keeps getting smarter — scheduled agents maintain it on a cadence, contradictions get reconciled instead of piling up, and I measure its recall with a small eval harness so search quality is a number, not a vibe. And this site: freshly rebuilt on Astro, with a build-log post drafting.
Learning
Mostly by building: agentic workflows — what it takes to let an AI operate a real system (my vault) safely, with audit trails and honest failure modes. The interesting problems are the boring ones: reconciliation, provenance, and knowing when the model is wrong. Reading more, too — the research on verifying and trusting software, and on making it work across languages.
Shipped recently
- Rebuilt this site on Astro + Tailwind — live at shadiul.github.io
- Embedded a live Mevrik chat widget — the one I build, running on the site
- Build-time OG images, mobile nav, RSS, dark + light themes
- First blog post drafted — a build-log of the second brain
- A retrieval-eval harness that scores how well the vault answers