KY

How Far Can You Go Without Knowing What You're Doing?

I knew tax wasn't "the one." But I couldn't just wait for the perfect idea. So I used it as a training ground - go as far as I can, extract every lesson, then recalibrate. Eight months later, here's what I actually built. (Hint: it wasn't a tax company.)

BUSINESS ENTREPRENEURSHIPTECHNOLOGY AIINSIGHTS LEARNINGS

Eight months ago I started building an AI tax company in Singapore.

I knew on day one it wasn't "the one."

I wasn't passionate about tax. I had no vision for compliance software. But I had momentum, a tax-genius brother, and zero better ideas.

So I made a cold decision: use tax as a paid training ground. Go as far as humanly possible, extract every hidden lesson, then recalibrate when the curve flattened.

I just didn't expect the curve to flatten with such a brutal mirror.

The Five-Phase Accident That Wasn't an Accident

Phase 1: Fear Inoculation (March)

Network School threw me into Balaji office hours with five minutes' notice, conferences, cold pitches. I stopped feeling fear. ✅

Phase 2: Technical Proof (April–June)

Twenty-plus iterations → multi-agent tax system hitting ~90% accuracy on the data we had. Built entirely through plain-English markdown files that Claude turned into production code. ✅

Phase 3: Strategic Validation (May)

Pitched Balaji. He handed me a perfect playbook: target family offices via their law firms, do 10 for free → references → escape velocity. ✅

Phase 4: Fundraising Test (June)

Raised a $1M crypto term sheet in a week. Turned it down because the vibe was wrong and (if I'm honest) because I was already looking for the exit ramp. ✅

Phase 5: Market Collision (July–October)

Automated outreach to 839 companies. 6 meetings. 1 real data set. Zero paying customers. Discovered the entire SME market is happily locked into two incumbents (SingTax, TaxPedited). Family offices won't share data with an unknown startup. Game over.

Every single target I set, I hit.

That turned out to be the problem.

The Real Reason I Stopped

I told people: "Crowded market, data-privacy moat, bad founder-market fit."

Those are all true, but they're the polite version.

The brutal version is simpler:

  1. I don't actually know if the tax product works in production (we only tested on a handful of cherry-picked cases).
  2. I never executed Balaji's strategy (contacted zero law firms).
  3. In 839 outreach attempts I asked for the sale exactly zero times.
  4. Every "no" I got felt like relief, not pain.

Translation: I was never all-in.

I was collecting badges (validation ✓ funding ✓ tech ✓) so I could tell myself I tried, then walk away clean.

That's not a market problem. That's a conviction problem.

The Only Thing I Actually Built

A capability, not a company.

I can now walk into any complex domain and build production-grade AI systems using nothing but natural-language markdown files.

Input → .md. Output → .md. No traditional code. Domain-agnostic.

Everything else—tax, Singapore, compliance—was just the substrate.

The Patterns I Will Break Next Time

  1. Action bias over strategy — When the hard path looked hard (law firms + family offices), I pivoted to the easy path (accounting firms) and quit when the easy path was saturated.

  2. Celebrating capability instead of customers — I got really good at hitting arbitrary targets and calling it progress.

  3. Using "training ground" as permission to never commit — Smart way to stay in perpetual warm-up.

  4. Mentally quitting months before officially quitting — By August I was already 100% on other ideas while still pretending to push the tax product.

These patterns killed the tax project. They will kill the next one too unless I burn them.

New Rule Set

Old way: Build → validate → sell → discover nobody wants it.

New way: Talk → understand pain → validate demand → THEN build.

Execution is easy for me. Target selection is the new constraint.

What's Next

I'm looking for the problem where:

  • I either am the customer, or
  • the data is freely available, or
  • I care so much I'm willing to cold-call 839 people and actually close them.

If you have a real, painful problem that needs complex systems built through plain English and AI, reach out. I can build almost anything now. I just refuse to build the wrong thing again.

The training is over. Time to pick a target worth hitting.

If you ever used "not the real thing" as deliberate practice for "the real thing," I'd love to hear your story.