A verifiable-AI company that began at a mixing desk.

The engineer

Fifteen years of getting the signal right

Tiago Merlini Ferrão has been on his own since he was fifteen. At eighteen he was in the Netherlands on almost nothing; then London, São Paulo, back to Lisbon — a music label built and left, a decade and a half as an audio engineer: front-of-house, mastering, sound design, broadcast mixing for Portuguese national television. A craft where you don't get to say the signal is clean — the meters either agree or they don't.

Then a quieter chapter. A hard personal period, and later a COVID layoff, gave him a long stretch of time. He used it to go deep — streaming, broadcast infrastructure, and then the thing that would reorganise everything: blockchain, and self-taught, AI-assisted code. The instinct for spotting a technology early was the same one he'd trained at the desk: don't trust the room, check the signal.

The instinct

A tool built for himself — dinamic.eth

It began as infrastructure he wanted to exist: dinamic.eth, an Ethereum-native kit — dynamic ENS identities, on-demand token-gated subnames, two-tier IPFS pinning, on-demand contract deploys, on-chain agents. Still live today, owned by a single wallet that would come to thread the whole story. Underneath the features was one conviction, carried over from broadcast: a system's proof should be part of its output — re-derivable by anyone — not a badge on a page.

The thesis

What you read is what you execute

On 11 June 2026 he filed WYRIWE — What You Read Is What You Execute — input-provenance for AI inference, assigned ERC-8299. That was the moment the boilerplate became a company-sized idea: don't trust the machine that produced a result — re-derive the result yourself, from public data, without owning the machine and without calling its servers.

Don't trust. Recompute.

It isn't the same promise as trusted hardware or zero-knowledge proofs. Trusted execution asks "can you trust the machine that ran this?" — and needs you to own or attest the silicon. Recomputation asks a different question — "can you re-derive this yourself?" — and that answer doesn't change with who owns the hardware. Vértice owns no GPUs, and it costs the proof nothing.

The open house

Built in the open, with a working group

He didn't build it closed, and he didn't build it alone. He and Damon Zwicker started it; a core of six became the originals, and the reference implementations moved into a neutral home — trustless-ai — where every repository ships clone-run-recompute and recomputability is the definition of membership. The founding group, in the order they arrived:

The group grew from there — early on Tracie Myers (a validation-network interface), Blockbird (bounded agent actions) and Panini (service discovery & escrow) — and many more since: giskard, Zexo, Ryan & Jin, and others. The rooms don't share a channel, so Tiago became the weaver — relaying between them, and always arriving with a running system, not just a paper. Build the thing first; open the standard second.

The company

Vértice Criativo — making the work legible

His last broadcast job was behind him — left over a disagreement he wasn't willing to swallow — and the projects were adding up. On 7 July 2026 he incorporated Vértice Criativo in Lisbon. The name of the discipline became the name of the company: a legible home for the work after a long stretch of being all-in on his own projects.

It runs on two faces. Vértice ships the commercial work — implementations, integration, the recompute-UI, the client relationship. The primitives underneath release CC0, in the open, through trustless-ai — because the open, recomputable stack is itself the credibility proof. Take the instance, never the definition — a rule he applies even to himself.

The house style

The bugs it hunts tell you what it is

Across a summer of building together, nearly every real defect the group found had the same shape: a check that verifies something next to what it claims to verify — and prints green doing it. A byte-integrity test standing in for two-party agreement. A signature covering every field except the one the logic used.

None were caught by re-reading. Each was caught by someone who didn't write it, or by real data arriving.

And it isn't a principle in a frame — it's a live practice. The group catches this shape in each other's systems, still: a review verdict that had quietly collapsed "found real evidence" and "couldn't tell" into one green-looking result surfaced in our own pipeline this very week, caught by another member. Agreement is not verification, and the only cure is to recompute.

Today

Live, in production, opening real doors

The stack is running, not proposed: agents that emit recomputable receipts, a public no-auth verify endpoint anyone can hit, post-quantum key binding anchored on Ethereum mainnet, and a self-verifying capability registry — hand it any entry and it re-derives it live. The primitives are moving through the Ethereum standards process.

And the open work has started opening institutional doors: a European supercomputing allocation (EuroHPC · Deucalion) for zero-knowledge inference proofs, a European public-money grant track, and an eIDAS-ready path that lets a recomputable AI action compose under a qualified seal a regulator or court accepts. From a personal ENS kit to standards, compute and regulation — in a little over a year.