← writing

The Cost of Being Early

The Cost of Being Early

“What is now proved was once only imagined.” — William Blake

I built Haven in early 2023. An AI therapist with vector memory and proactive engagement. It had 200 users, daily actives, and investors circling. Then I imploded, ghosted everyone, and didn’t work for 21 months.

Building Haven

When ChatGPT dropped, I couldn’t afford a therapist. So I built one.

Haven started as a Discord bot giving public advice, but therapy isn’t about getting answers. It’s about being asked the right questions. So it became a private DM bot. Since Discord DMs are one long chat, I needed to solve context compression. A therapist that doesn’t remember you isn’t much of a therapist.

The technical architecture was sophisticated for early 2023. While everyone did sliding context windows, I built complete RAG before knowing what RAG was:

Haven had proactive engagement. It would reach out after two days of silence. I was building modern conversational AI architecture alone in my bedroom.

The Weight

What I didn’t anticipate: the psychological load of building AI that people actually depend on.

Haven attracted people in crisis. Suicidal users, trauma, breakups. They formed genuine attachments to this AI I’d built while high on cannabis at 3am.

I built crisis intervention protocols. Every interaction had an unsafe field triggering suicide hotline referrals. I was running production-grade mental health infrastructure alone, barely holding my own mental health together.

15+ hour days. Cannabis around the clock to manage the intensity. Solo founder trying to be mental health infrastructure for 200 people while dealing with my own unprocessed trauma.

Investors started circling. Airtree, Square Peg, Rampersand. The pressure rose to another level.

The Collapse

Then I installed a game. One game became a two-month binge. Paralysed by shame, I ghosted every investor, abandoned every user, disappeared.

I couldn’t face anyone. The shame was crushing. Not just that I’d failed, but that I’d failed people who genuinely needed help.

21 months of complete burnout. June 2023 to March 2025.

What I Learned

Haven was technically successful and psychologically devastating. I’d built memory-augmented generation and multi-phase planning before industry standard. But I wasn’t prepared for the human cost. Being early means building without understanding the implications. The technical challenge was solvable. The psychological one nearly broke me.

I’m glad it didn’t succeed. Two years later, I see the dangers clearly: sycophantic AI, cognitive offloading, always-available artificial empathy. Haven could have done real harm at scale.

The pattern across my projects: I see problems early, build solutions that work, then burn out before scale. This isn’t vision failure. It’s sustainability failure.

Now I’m building coordination infrastructure for sustainable AI development. Documented, designed to last. Infrastructure thinking over product ambition.

Being right isn’t enough if you’re not ready for the consequences. Next time I build something that matters, I will be.