All Articles

How I Built My Portfolio: A Headless CMS, a Free AI Assistant and Lessons in Speed

A Next.js portfolio with its own CMS, a free AI assistant that never invents facts, recruiter tools, and what I learned making it fast.

Nilay NahaNilay Naha
•Oct 28, 2026•7 min read•7 views
How I Built My Portfolio: A Headless CMS, a Free AI Assistant and Lessons in Speed

Most developer portfolios are a static page with a few project cards. I wanted mine to work like a small product: a site I can update from my phone, that answers recruiters' questions on its own, and that costs me ₹0 a month to run. This is how I built it, what's inside, and what I learned along the way.

Next.js TypeScript MongoDB Headless CMS Web Performance

₹0Monthly running cost
0Paid AI API keys
286Skills in the tech catalog
The home page
The home page

Why I built my own instead of using a template

Templates are fast to start with, but every content change means editing code and redeploying. I wanted three things a template couldn't give me:

  • Edit everything without touching code: projects, blog posts, my bio, experience, testimonials and even the home page headline.
  • Help recruiters fast: let them ask a question, check a job fit, or grab a résumé in one click.
  • Stay free: no paid hosting, no paid storage, no paid AI API.

The tech stack

LayerWhat I usedWhy
FrameworkNext.js 14 (App Router)Server rendering, ISR and API routes in one project
LanguageTypeScript (strict mode)Catches mistakes before they reach production
StylingTailwind CSS + Framer MotionFast to build, smooth animations, light and dark themes
DatabaseMongoDB Atlas + MongooseFree tier, flexible schemas for content
HostingVercelFree deploys on every push
MediaGitHub assets repo + Google DriveFree image and video hosting

A custom headless CMS

Behind /admin there's a full content management panel I built myself. From there I can create and edit projects (with case-study fields like problem, approach, results and metrics), blog posts, the About page, testimonials, the home hero, and the chatbot's curated answers.

A few details I'm proud of:

  • Drafts are never public. A post stays hidden until I publish it. Even when I'm logged in, the public site shows exactly what visitors see.
  • Rich content without uploads. Posts support Markdown, KaTeX math, code highlighting, design classes like callouts and stat cards, and Google Drive images, GIFs and videos. I paste a Drive link and the site turns it into a working embed.
  • Changes go live instantly. Every save clears the cached pages and data, so I never wait for a rebuild.
Free media, on purposeI don't pay for storage. Binary files like my résumé go to a separate GitHub repository that acts as a free CDN, and in-article media comes from Google Drive. The main code repository stays small.
The projects page, filled entirely from the CMS
The projects page, filled entirely from the CMS

An AI assistant that costs nothing

The chatbot in the corner answers questions about my projects, experience, skills and education. It does not call ChatGPT or any paid model.

Instead, it's a small answer engine I wrote:

  • It reads my live CMS content, so its answers update the moment I edit something.
  • It ranks content with BM25 full-text search plus intent detection ("what projects…", "does he know…", "tell me about him").
  • It never invents facts. If something isn't on my portfolio, it says so honestly and points to the contact page.

It also has voice input, read-aloud answers, nine languages, and Claude-style loading words ("Pondering…", "Rummaging through projects…") while it thinks. Ask it for my résumé and you get it instantly.

Why not a real LLM?An LLM API costs money per message, can stop working when a key runs out, and can confidently make things up about me. For a portfolio, a search engine over my real content is cheaper, faster and more honest.

Built for recruiters

I added tools for the people who actually read portfolios:

  • Job match (/match): paste a job description and get a match score with evidence from my projects and skills.
  • Tailored résumé (/resume/tailored): a one-page résumé generated from my live portfolio data, reordered for a specific role.
  • Command palette: press Ctrl + K to search every page, project and post.
  • Book a call and Get notified push notifications for new posts and projects.
The job-fit match page (/match)
The job-fit match page (/match)

Discoverable by search engines and AI

More people now find developers through AI assistants, not just Google. So the site exposes its content in machine-readable ways:

  • JSON-LD structured data, a sitemap and robots.txt rules for AI crawlers.
  • An llms.txt file summarising who I am for language models.
  • A public read-only MCP server, so AI agents can query my projects and skills directly.
  • OpenAPI docs for every API route at /api-docs.
Interactive API docs at /api-docs
Interactive API docs at /api-docs

Security I didn't skip

A personal site still has an admin panel, so I treated it seriously:

  • Admin sessions use signed, expiring cookies. No secret is ever stored in the browser.
  • Logins are throttled: after 5 failed attempts you wait 15 minutes.
  • Every public endpoint is rate-limited, and inputs are sanitised against NoSQL injection.
  • Security headers and a Content Security Policy restrict what the page can load.
  • Every admin change is written to an audit log.

Making it fast: lessons from this week

This is the part that taught me the most. Visitors told me the site felt slow, so I dug in.

  1. Problem 1: every visit hit the database. The public APIs were set to never cache, so each new visitor waited on a MongoDB query. I moved them to Next.js's server data cache, tagged so any CMS save clears them instantly.
  2. Problem 2: lists carried full articles. The blog list sent the complete text of every post just to show cards. Now it sends only what the cards need.
  3. Problem 3: 3D on every page. I had a 3D chatbot mascot (three.js plus a 266 KB model) that loaded on every page. On phones it was worse: each mascot opened its own WebGL context, and mobile browsers silently kill the oldest ones, which showed as broken images.
  4. The fix: I removed the 3D mascot completely and replaced it with a single 15 KB image. The remaining WebGL effect, a cursor trail, now loads only after the page is idle and never on touch devices.
The honest lessonThe 3D mascot was the feature I was most excited about, and it was the one hurting the experience most. A fast page beats a fancy one.
A blog article with a sticky table of contents
A blog article with a sticky table of contents

Other things that make it feel like an app

  • Installable PWA that works like an app on your phone.
  • Swipe navigation between pages on mobile, tuned so vertical scrolling never triggers it by accident.
  • Reload memory: refresh and you land back where you were.
  • User-flow analytics I can read in the admin panel, without a third-party tracker.
  • A chat window that's full-screen on phones and floats on tablets, laptops and TVs.

Key takeaways

  • Build the CMS early. Being able to edit from my phone means I actually keep the site up to date.
  • Free can be good. MongoDB's free tier, Vercel, GitHub and Google Drive cover everything a portfolio needs.
  • You don't always need an LLM. Good search over real content gives honest answers for free.
  • Cache first, decorate later. Most slowness came from uncached data and heavy visuals, not from the framework.
  • Test on a real phone. Two of the bugs I fixed this week only appeared on mobile.

The portfolio keeps evolving, and this post was itself published through its own admin tools. If you'd like to see how any part works, ask the chatbot or get in touch through the contact page.

Topics Covered
#Web Development#Backend & APIs#Next.js#TypeScript#MongoDB#Headless CMS#Web Performance#Portfolio
Explore More Articles