Every side project I finish looks obvious in hindsight. This one took two dead ends to find its shape, and the dead ends taught me more than the final idea did.

The product is called Deckname. It is a pseudonymization gateway for small law firms. You paste text or a PDF, it detects the personal data, you review it, and you get back a version where each value is a placeholder like [PERSON_1]. You send that to a language model, and when the answer comes back, Deckname reverses the placeholders. The model never sees a real name. It runs entirely on the firm’s own hardware.

The short version of how it got here: I wanted to build something at the intersection of privacy and AI, guessed wrong about the shape twice, and only found the market once I stopped trying to hold other people’s data at all.

Dead end one: a server that cannot read

The first idea was an encrypted knowledge base for AI agents. Think of a hosted notes app with zero-knowledge storage: your notes live on my server, encrypted so I cannot read them.

It died on one contradiction. The feature people wanted was retrieval: ask a question, get the relevant notes back, feed them to a model. Retrieval means comparing a query against content. If the server cannot read the content, it cannot compare. A server that cannot read cannot search.

I worked through three escapes, and each one gave the premise away:

  • Do the retrieval on the client, where the plaintext lives. Then the hosted service barely does anything.
  • Decrypt on the server only while the user is active. Now the server sees plaintext, just briefly, and the promise is gone the moment it matters.
  • Keep the notes encrypted but store the search embeddings in the clear. Embeddings leak. You can reconstruct a lot from them, so the “clear” part is the sensitive part.

The honest conclusion: if a server has to understand your content to be useful, “it cannot read your content” is not a promise you can keep.

Dead end two: a market I could not enter

Second attempt: privacy-preserving AI tools for German public authorities. The research ended it fast. Procurement law, ISO 42001 turning up in tenders, BSI C5 certification, and the state already building its own platforms. As a solo project on evenings and weekends, that is not a market you walk into.

But the same demand lived right next door, with people who can actually decide to buy. In Germany, professionals bound by § 203 StGB (lawyers, doctors, tax advisors, notaries, psychotherapists) hold data they are legally required to protect. A small law firm does not run a tender. It buys a tool that solves a problem this month.

The niche, and the word that defines it

So the shape settled: a pseudonymization gateway for small firms, roughly two to ten people, UI first, on-prem. Not a notes app. Not a public-sector platform. A narrow tool for a group with a sharp, unmet need.

The market taught me one more thing: be honest about what the tool does. The enterprise options are priced for enterprises. Some consumer tools lean on bold accuracy numbers. One product got publicly taken apart for not living up to its “anonymization” promise. That teardown was the lesson. In this field, an honest “pseudonymization” beats a confident “anonymization” that does not hold.

So Deckname says pseudonymization, never anonymization. Pseudonymization is reversible by design, which is the entire point of the reverse step. There is no “100%” and no “guaranteed” anywhere in the product. Detection is assistive. The human review step is part of the product, not a disclaimer at the bottom.

The name carries all of that. “Deckname” is the German word for a cover name. It describes exactly what the tool does, which is why it won over the alternatives.

Where the two dead ends led

The failures were not detours. They pointed straight at the design. Deckname does not try to hold your data and stay blind to it. It refuses to hold it at all. The plaintext never touches a server I run. The reversible mapping is encrypted at rest, and the one component that could read it runs on the customer’s hardware, not mine.

The first idea was right about the goal and wrong about the shape. It took building the wrong thing twice to see the right one.