andr3s.com

An AI-powered portfolio website — ChatGPT, but it only knows about Andres.

Recruiters and hiring managers can chat with an AI agent that explains my background, projects, and experience — all grounded in a curated knowledge base.

This project was also an experiment.

I wanted to see how far modern AI coding tools could go if I let them take the lead.

~90% of this repository was written by Claude Code.

My role was mostly:

• defining the architecture
• reviewing and steering the generated code
• integrating it with systems I'm already comfortable with (like LiveKit agents)

The result surprised me — both in speed and quality.

I'm intentionally leaving this repo public as a demonstration of what modern AI-assisted development looks like in practice.

---

Architecture

The site is essentially a chat interface backed by an AI agent that answers questions about me.

Frontend

• React / Next.js
• ChatGPT-style streaming UI
• Markdown-rendered responses

Backend

• Lightweight PHP API layer
• Handles requests from the frontend
• Communicates with the LLM
• Protects API keys from the browser

I chose PHP intentionally because the site is deployed on a very cheap Hostinger shared hosting plan, which limits what runtimes can run server-side.

The goal was to prove you can still build something modern even within those constraints.

Knowledge Base

The AI does not know anything about me unless it's written in the repo.

All knowledge lives in: knowledgebase/

These Markdown files are embedded into a vector database that the agent queries before answering questions.

This ensures responses are grounded in real content instead of relying on model memory or hallucination.

Updating the AI's knowledge is as simple as editing Markdown.

---

LiveKit Integration

The chat agent runs through LiveKit agents, which provide:

• real-time streaming responses
• event-driven AI workflows
• a clean abstraction around LLM execution

This is the part of the stack I'm personally most familiar with, so integrating it with the AI-generated codebase was straightforward.

---

Deployment

This site is intentionally hosted on a very cheap Hostinger plan.

Because of those constraints:

• backend logic runs in PHP
• the frontend is a built React app
• deployment uses GitHub Actions + SSH + rsync

Every push to main automatically:

1. builds the React app
2. uploads the build to the server
3. syncs the PHP backend

You can see the deployment pipeline here: .github/workflows/

---

Why this repo is public

Two reasons:

Demonstrate modern AI-assisted development

Tools like Claude Code dramatically change how quickly small projects can be built.

This repo is a real-world example where:

• AI wrote the majority of the code
• I provided architecture, direction, and review
• the end result is production-ready

Show how simple a personal AI agent can be

You don't need a massive stack to build something like this.

The entire project runs on:

• Markdown files
• a vector database
• an OpenAI-compatible LLM
• a cheap shared hosting plan

---

Running locally

Front end:

npm install
cp .env.example .env
npm run dev

Back end:

php -S localhost:8080 router.p