Tidbix

• 

Inspiration

Most founders fail by building for nine months before finding out they were wrong. The opposite failure is just as common and less discussed: going back to the drawing board every time the world moves. Same root cause — no moment where you are actually allowed to decide.

I have run the second failure mode myself. Tidbix builds that missing moment into the structure and makes you bring evidence to it.

What it does

Tidbix is a venture pathway. It takes a founder from purpose to paying customers across five phases — Mission Foundation, Venture Idea Lab, Evidence Sprint, Build Studio, Launch & Grow. Each phase runs on numbered steps, and each step ends somewhere most founder tools never go: a stage gate with four doors. Proceed. Revise. Pause. Stop.

Stopping with evidence counts as a successful outcome here. That is the opposite of what almost every other founder tool communicates.

Where Tidbix actually is

Free access is live at www.tidbix.com. Paid access has not opened, and nothing has ever been sold — zero revenue, zero arms-length users. Thirteen profiles exist in production and every one is the founder or a tester: related-party, disclosed as such, and they should be removed from consideration as customer traction. Marketing and customer-acquisition spend is $0.

That is verifiable rather than merely stated. Stripe is integrated and deployed, and the entitlement ledger shows why the zero is real: all 23 access grants in production are free defaults, comps, or manual grants, and not one carries a Stripe reference. Every payment event on record is a comped grant. The payment rail exists, is wired, and has never fired.

I am stating that first, unprompted, because it is the only position consistent with the product. Tidbix exists to stop founders building on untested assumptions. A Tidbix submission that claimed traction it did not have would be refuted by its own thesis. Business Viability is one of three judging criteria and this entry will score near zero on it. That is a real cost, and I would rather accept it than dress it up.

How we built it

Twenty-two Supabase edge functions are deployed and active in production. Eighteen of them route decisions through the Gemini API — gemini-3.7-flash, called via generativelanguage.googleapis.com, billed through the Google Cloud project Gemini XPrize. The remaining four are deliberately deterministic.

The eighteen are not a chat wrapper. They read what the founder has already recorded and produce structured, validated output. idea-fit scores an idea across six dimensions against that founder's own mission, energy map and values rather than a generic rubric. audience-review reads a stated audience, tells the founder when it is too broad, and names the arena it thinks they are really in. founder-council pressure-tests a venture, leads with where the pattern breaks, carries its sources, and says so when two lenses disagree instead of picking one.

Three architectural commitments make that safe to ship:

Counts are computed in code, never asked of the model. Anything the product depends on numerically is derived deterministically. The model is asked for judgment, not arithmetic.

Every AI response passes a fail-closed validator before it renders. If the response does not match its contract, nothing ships to the founder — no partial render, no degraded output.

Fields that record what happened in the world get no AI suggestion at all. An interview, an outreach attempt, a piece of evidence — there is no suggest button, and there never will be.

Challenges we ran into

The hardest call was one we chose not to make. Embeddings remain on OpenAI's text-embedding-3-small while chat runs on Gemini. The pgvector columns are 1536-dimensional and were built by that model; mixing embedding spaces does not error, it silently returns wrong matches. Migrating embeddings would mean re-embedding every stored vector. A silent failure mode is worse than a mixed stack, so the mix stayed. That is a considered decision, not a gap, and disclosing it beats quietly misstating the architecture.

The second was discipline about timing. AI has executed decisions in Tidbix production since June. What changed on August 14 was the LLM provider — chat moved to the Gemini API. The commit and the Google Cloud billing record corroborate each other, and both are available to judges.

Accomplishments that we're proud of

The feature we deliberately did not build. No AI suggestion appears on any field that records what actually happened in the world. It would have been trivial to add and it would have made the product demo better. It is absent on purpose, because the moment a tool will draft your customer interview for you, the evidence stops being evidence.

The AI can help you think. It cannot help you pretend.

The human/AI labor split

Tidbix is one person. All 185 commits are authored by the same account, between June 6 and August 17, 2026 — about 90 days.

Human. The vision and the intent are mine: what Tidbix is for, the five-phase pathway, the stage gate with four doors, and — most consequentially — the decisions about what the AI is not permitted to do. Each of the three commitments above constrains the system rather than extending it, and each was a product decision, not an implementation detail. I also set the copy and positioning constraints the codebase enforces in CI. I shaped that thinking in conversation with Gemini, ChatGPT and Claude; the judgment calls and the accountability are mine.

AI. Claude's Fable 5 model wrote the majority of the backend systems now running in production — the edge functions, the validators, the data layer. Gemini and ChatGPT worked alongside Claude through the thinking and drafting stages. Every AI-written change was reviewed before merge.

Stated plainly: a solo founder set the intent and the constraints; AI wrote most of the code that runs. Those three commitments exist precisely because building this way at this speed needed guarantees neither the assistant nor the founder could quietly skip.

What we learned

Building this fast with AI does not remove the need for judgment — it concentrates it. The scarce resource stopped being implementation time and became knowing which constraints to impose before the code existed to violate them.

What's next for Tidbix

Users and revenue are the next ninety days. I will test that the same way I am asking every founder to. With evidence.

Built With

• ai-native
• build-with-gemini
• deno
• entrepreneurship
• evidence-based-decisions
• founder-tools
• gemini-api
• google-cloud
• next.js
• openai
• pgvector
• postgresql
• posthog
• react
• resend
• startup-validation
• stripe
• supabase
• supabase-edge-functions
• tailwindcss
• typescript
• venture-building
• vercel
• vitest
• xprize