2
5 Comments

I built a production SaaS in 100 hours by letting AI agents write all the code

I'm a solo founder. No co-founder, no dev team, no funding. Just me, a Linux box in my house, and a pipeline of AI agents.

The backstory: I spent a few months building AI prototypes at a stealth-mode startup. The engineering was interesting but the company couldn't ship. I kept watching the AI tools I was experimenting with get better every week — and I realized they were more valuable than what I was building with them.

So I left and started Noncelogic. December 15th, 2025.

What I built: Passband (passband.ai) — an AI content intelligence platform. It monitors RSS feeds, Hacker News, Reddit, and arXiv for industry news. It scores everything for relevance using AI. Then it drafts social media posts in your calibrated voice. You review and approve before anything publishes.

Think of it as: Feedly's brain + Jasper's drafting + Buffer's scheduling — except it actually sounds like you, not ChatGPT.

How I built it: I didn't write the code. I wrote the tickets.

I set up an event-driven pipeline where:

  1. I write a GitHub issue with clear acceptance criteria
  2. An orchestrator bot picks the highest-priority unblocked ticket
  3. Claude Code implements it in a sandboxed branch
  4. CI runs (build + 12 Playwright E2E tests)
  5. If green, I review and squash-merge. If red, it retries with error context.

WIP limit: 1 ticket at a time. Average time from ticket to merged PR: ~45 min for small tasks, 2-3 hours for medium.

The numbers (100 hours, Dec 15 – Feb 14):

  • 30+ tickets shipped
  • 12 E2E tests passing in CI
  • 2 live publishing integrations (Bluesky, Mastodon)
  • Full pipeline: ingestion → AI scoring → voice-calibrated drafting → scheduling → auto-publishing
  • Stripe billing wired up
  • AI chat interface with 14 tools
  • Deployed on Vercel, DB on Neon Postgres

I worked 10 hours/week max. The AI agents worked 24/7.

The model that makes it sustainable: BYOK — Bring Your Own Key. Users plug in their own OpenAI or Anthropic API key. A full pipeline run costs ~$0.04 in their tokens. I charge for the platform, not the compute. This means I can sell lifetime deals without going bankrupt on inference costs.

What's next:

  • Landing page for passband.ai
  • Lifetime offer launch (3-tier, direct — not AppSumo)
  • Multi-profile support so agencies can manage multiple client voices
  • LinkedIn integration (once I get API access)

I'm targeting professionals who need thought leadership content but don't have time to write it — fractional CMOs, financial advisors, solo lawyers. People whose reputation depends on sounding like themselves, not a bot.

Happy to answer questions about the architecture, the AI pipeline, or the business model. Also — if you're in the NYC area and want to grab coffee, I'm always down.

on February 21, 2026
  1. 1

    Hi Joe, your project caught my attention because I'm currently learning backend development and trying to understand how real production systems are built. I'm curious about your experience using AI agents to write the code — what parts of the system still required the most human engineering decisions?

  2. 1

    100 hours is impressive. Curious which model handled the hardest parts. In my experience the deploy loops eat more time than the actual coding.

  3. 1

    That’s a wild build loop and cool results. One thing we noticed after forcing things into prod was that integration and reliability still needed orchestration outside the core agent pipeline. Anyone else see the “AI writes code fast, but glue is the real work”?

  4. 1

    dashboard page is broken

    1. 1

      thank you for reporting this, you have been given "Pro" tier for your kindness. it turns out my DB migration pipeline was missing a secret and it didn't run. thanks so much again for your help.