Why the same tools speed up experienced teams and quietly break apps built without them. Notes from the team behind Walletverse
We write this newsletter as the Walletverse team. Walletverse is a self-custody crypto wallet made by ilink, a company that develops fintech and crypto software. A wallet is the kind of product where the debate about AI-written code stops being abstract: the code handles private keys and builds transactions that cannot be reversed. So we often hear a fair question: do you use AI to build it?
Yes. How we use it is the subject of this post.
"Reviewed and accepted"
In October 2024, Sundar Pichai wrote that "more than a quarter of all new code at Google is generated by AI, then reviewed and accepted by engineers." The first half of that sentence travelled widely. The second half travelled less.
Reviewed and accepted by engineers: those words mark the difference between two ways of using the same tools. In one, AI drafts code and a person who understands the system decides whether it goes in. In the other, AI drafts code and it goes in because the app appears to work. The second is what Andrej Karpathy called vibe coding in February 2025, and he described it as "not too bad for throwaway weekend projects." The caveat got lost as the phrase spread.
An amplifier, not a replacement
Google's DORA research program surveyed nearly 5,000 technology professionals for its 2025 report, and 90% of them said they use AI at work. The report's central finding fits in one sentence: "AI doesn't fix a team; it amplifies what's already there." AI adoption was linked to higher delivery throughput and better product performance, but also to less stable delivery, and the authors name what makes the difference: "Without robust control systems, like strong automated testing, mature version control practices, and fast feedback loops, an increase in change volume leads to instability."
Read this way, the wave of leaky AI-built apps is not a story about AI being bad at code. When security researchers scan public vibe-coded apps, they keep finding the same things: missing access rules, keys shipped to the browser, logins that can be bypassed. That is what amplification looks like when the safeguards professional teams rely on were never part of the process.
"Almost right, but not quite"
Professional developers are not AI skeptics. In the 2025 Stack Overflow Developer Survey, 84% said they use AI tools or plan to. Yet 46% said they distrust the accuracy of AI output, against 33% who trust it. The most common frustration, named by 66%, was "AI solutions that are almost right, but not quite." And 72% said vibe coding is not part of their professional work.
These numbers fit together. Developers use AI every day and still do not take its output on faith, because "almost right" is the most expensive kind of wrong. Code that fails loudly gets fixed. Code that works in the demo and mishandles an edge case, such as a rounding rule, an expired session or a transaction submitted twice, goes to production. Catching it requires knowing what right looks like.
What vibe engineering looks like
In October 2025, Simon Willison, a long-time developer and co-creator of the Django web framework, proposed a name for the opposite of vibe coding: vibe engineering. He describes experienced professionals using coding agents to move faster while staying fully accountable for what they ship. His list of what makes agents useful reads like a description of a senior engineer: automated tests, a plan agreed before any code is written, good documentation, disciplined version control, a code review culture, careful manual QA, and the judgment to decide what to hand to an agent and what to do yourself.
"AI tools amplify existing expertise," he writes. "The more skills and experience you have as a software engineer the faster and better the results you can get."
That is also the answer to a question people often type into Google: what is the difference between vibe coding and vibe engineering? Not the tools. The process around them, and the person who answers for the result.
How we use it at ilink
At ilink, the company behind Walletverse, AI tools take on routine work. Decisions stay with engineers and developers. Team leads review the code, and QA engineers and testers check everything. We always control what the AI does. It does make us faster. We just do not let speed decide what goes into a product that holds people's money.
For a wallet this is not a matter of style. Walletverse is self-custodial: private keys stay with the user, and, as we put it on X, "We can't access your wallet, approve transactions, or lock your funds." That promise lives in code: in how keys are generated, encrypted and stored, and in how a transaction is put together before the user signs it. The Walletverse case study on the ilink site describes five security and trust layers, including AES, ECDSA, BIP32/39/44, MFA and biometric access, along with encrypted key management and third-party smart contract and security audits. None of that is code you accept because the screen looks right.
If you don't have a technical background
None of this is an argument against non-engineers using AI. A prototype built over a weekend is a great way to test an idea, show it to users and give a development team a clear brief. Many good products will start exactly that way.
The line is somewhere else: real users, real money and personal data. Past that line, someone needs to read the code, own the access rules and the keys, and be able to explain why the product is safe. If that person is not on your team yet, it is better to find them before launch than after. This is part of what we do at ilink as well: we take prototypes built with AI or by another contractor, review the code and bring them to production. More about our approach to crypto wallet development.
Same AI, different hands. The tools will keep getting better. The difference will stay with the people using them.
Top comments (0)