Building a Full-Stack AI Portfolio OS in 25 Days

The Challenge

A portfolio site is supposed to prove you can do the work. For a developer building AI systems and full-stack platforms, that means showing — not telling. The challenge: build a production portfolio that itself demonstrates autonomous AI agents, programmatic SEO, full-stack admin panels, and modern dev automation.

Time constraint: 25 days. No team. Just Claude Code, Next.js, and a clear vision.

The Stack

• —Next.js 15 with React 19, TypeScript, App Router
• —Supabase for PostgreSQL, auth, RLS policies, and real-time
• —Vercel for deployment, CI/CD, and ISR edge caching
• —Claude API for the autonomous agent system
• —Groq (Llama 3.3 70B) for fast agent inference
• —DeepSeek R1 for complex reasoning tasks

The Build

Phase 1: Core Platform (Days 1–8)

Built the public portfolio with blog, services, case studies, and contact. Implemented ISR (revalidate: 300). Set up Supabase schema with RLS policies. Deployed to Vercel on day 3.

Phase 2: Admin CRM (Days 8–14)

Full admin panel with Supabase auth, client pipeline (lead → prospect → proposal → active), task management, content planner, and email subscriber management. Mobile-first with bottom tab navigation.

Phase 3: Agent System (Days 14–20)

Autonomous Claude API agents with tool-use loops. Agents read business context, propose blog drafts and client tasks, log to pending_actions, and wait for human approval. Never write directly to production tables. Daily brief runs via Vercel Cron at 9am weekdays.

Phase 4: Programmatic SEO + Content (Days 20–25)

60+ industry pages and 20+ local pages using generateStaticParams. 24 blog posts seeded via SQL with dollar-quoted markdown. Schema markup on every page. llms.txt for AI search crawlability.

Results

Four production platforms shipped: the portfolio OS, the news platform (TheWestNepal.live), the AI agent marketplace (agenticai01.tech), and a US gaming client backend — all within the 25-day sprint.

The AI agent system runs the daily brief autonomously, proposes content, and surfaces client action items — all reviewable from the admin panel. Zero rogue actions since launch.

Key Learnings

Claude Code compressed development time by ~60% versus traditional coding. The approval workflow pattern (agents propose, humans approve) is the right mental model for autonomous AI in production. Supabase + Next.js + Vercel is the modern founder stack — zero infra ops, scales to millions.