Building 10 AI Startups in Parallel: What Year One Taught Me

I committed to building 10 AI products in parallel at the start of 2024. By April 2026, six are live in beta, four are still being built, and two of the original ten got cut and replaced. This post is the honest version of what year one looked like, what survived contact with reality, and why I'm still doing it instead of focusing on a single bet.

The short version: parallel product development with a small senior team is harder than it looks, but the compounding learnings across products are real, and the alternative — betting everything on one product I'm wrong about — feels worse than the operational complexity of running ten.

Why ten, not one

Most founders building AI products are committing to a single thesis. The thesis might be right. It's also unverifiable at the time of commitment. The data we have on what AI products will sustain pricing power, what becomes commoditized by next quarter's model release, and what stays valuable enough to anchor a business — all of that is being written right now.

Ten products is a portfolio bet on uncertainty. Each one is small enough to kill cheaply if it doesn't work, large enough to learn something durable, and chosen specifically to share infrastructure with the others so the marginal cost of the eleventh is low.

What survived the first year

• Eval harness discipline. Every product that's still alive has a measurable definition of 'good' that we set before model selection.
• Shared infrastructure. The auth layer, the analytics, the deployment pipeline — same code, same patterns, across all ten.
• Internal-first deployment. Every product that worked was used inside NextBangla for at least four weeks before opening to external users.

More to come — this post will keep growing as the year-two retrospective lands. The full case studies for Voxly and NOBBYO are in the portfolio if you want the technical detail.