The site you're reading this on was designed, built, tested, and shipped using the same AI-assisted workflow described below — not a separate demo of the method, the actual production artifact.
A portfolio has to do two things most portfolios only claim to do: prove real, domain-rooted engineering experience, and prove that AI-assisted development actually produces trustworthy, maintainable, shipped work — not a one-shot demo that breaks the moment someone asks a follow-up question. Telling a visitor "I can build with AI" is a claim. This page is the evidence.
Anyone evaluating whether to hire me — for a role or a freelance build — who has reasonable doubt about what "AI-assisted" actually means in practice: a person who typed one prompt and shipped whatever came out, or someone who runs a real requirement-to-ship workflow with a human making every judgment call.
Static site, no framework, no build step — a single-page portfolio doesn't need one. Every factual claim on the site has to be traceable to a real repo or real coursework — no invented metrics. Positioning and tone rules had to survive every edit, including three deliberate title reversals over six weeks, without silent drift.
A fast, accessible, installable static site with dark/light theming, interactive click-to-explain diagrams, dedicated case-study pages per project, and a single source of truth for every fact that also has to appear, worded consistently, in two generated CVs.
Select a stage to see what it does.
Vanilla HTML/CSS/JS over a framework — one page doesn't justify build tooling. A single project-level instructions file acts as the standing "constitution" every AI-assisted change is checked against — including hard content rules like "every skill claim must trace to a real repo or real coursework" — so the honesty constraint travels with the project, not just with one session's memory.
A requirement comes in — in English or mixed Bangla/English — Claude Code reads the relevant files, drafts the change matching the existing patterns (BEM-ish class names, the same pipeline diagram component used across sections), and I review every diff before it's applied.
Before any push: grep checks for banned terms and policy violations, an HTML tag-balance check, a manual click-through of navigation and both CV downloads, and — when content changes — regenerating both CVs and confirming with LibreOffice that they still render at exactly two pages.
Cloudflare Pages, deploying straight from the main
branch — no separate staging environment for a static site this
size; every commit that lands is reviewed first.
A production site that has survived real iteration, not a frozen demo: three deliberate positioning reversals, a full section redesign, and a growing set of case-study pages like this one — all shipped through the same workflow, all still live and traceable in git history.
AI speeds up the mechanical part — drafting markup, keeping facts in sync across the site and two CVs, catching a missed file. It does not decide what's true, what's worth claiming, or when a positioning change is warranted — those stayed human decisions on every single change, which is the actual point this page is trying to prove.