Solvd learned to
break software.
Now it builds AI.

The chatbot existed. That was the problem. At a large American credit union, a previous team had built an assistant that never reached production. It could not connect to core banking. It had accessibility gaps. Members still needed people to answer routine questions. A piece of software had been created; a useful service had not. In Solvd’s account of the rescue, the distance between those two things is where the work began.

First, make it survive Monday.

Solvd rebuilt the credit union’s assistant around better intent recognition, secure authentication and connections to account information. Capabilities went live incrementally after review. The company says the system now operates across web and mobile, with less reliance on a third-party live-chat platform. The telling detail is the order: understanding the question, gaining permission, reaching the data, then answering. Fluency alone would have left the original problem intact.

That makes this a useful opening to Solvd, an engineering services company whose current pitch revolves around AI in production. Its customers buy help with systems that have to fit an existing business: applications, data pipelines, cloud infrastructure and testing. The work reaches product companies and enterprises across finance, retail, healthcare and technology. Named customers in its published cases include NerdWallet and MyFitnessPal. An anonymous banking project offers a particularly clear explanation of the job.

The advantage of having been the tester

Founded in California in 2011, Solvd started with quality assurance and test automation. Co-founder Alexey Khursevich’s public professional profile describes that engineering background. Testing is an unusual apprenticeship for an AI pitch: the daily task is finding the situation in which apparently functioning software stops behaving. A successful demonstration is only the beginning of the examination.

Consider MyFitnessPal. Solvd introduced Carina, its automation framework, to share test code across iOS and Android. A mobile cloud ran tests on 30 devices around the clock. Solvd reports 80 percent reuse between the platforms’ automated test cases and 35 percent lower QA costs. Those are supplier-reported results, but the mechanism is legible: write less duplicate testing code, run it repeatedly, catch defects before a release.

At NerdWallet, the machinery was more literal: a private mobile cloud containing 36 iOS devices and 38 Android devices. Strict security protocols shaped the setup, while simulated financial interactions kept testing away from production changes. Solvd reports 95 percent of test cases automated and 90 percent regression coverage. The underlying lesson is pleasantly concrete. Sometimes better software requires a cupboard full of phones.

The company also built a route into the profession. Solvd.Laba, launched in 2014, offered free hands-on training in QA, test automation and JavaScript. That is a specific expression of its stated interest in learning and collaboration. For a services business, developing engineers is part of developing the product: the people assigned to a client are the people who must understand the awkward exception.

Then buy the missing pieces

The company’s expansion has a financial spine. In September 2022, Siguler Guff took a majority stake, backing international delivery and a broader engineering bench. In October 2024, Solvd acquired EastBanc Technologies, adding AI consulting, data and infrastructure expertise. That announcement explicitly described a pivot toward the AI opportunity. The shift was a stated commercial decision, rather than a conversion story assembled after the fact.

In June 2025 came Tooploox, the Polish AI development and research company. The acquisition announcement described a roughly 200-person team and more than 50 published research papers. Consulting helps identify what a client should attempt; applied research helps with problems that resist an ordinary implementation. Bringing both alongside software delivery is Solvd’s answer to a buyer who would otherwise coordinate several suppliers.

There is a tangible research credential, too. A paper co-authored by Chief Scientist Tomasz Trzcinski received a NeurIPS 2025 Best Paper Award for work on deep self-supervised reinforcement learning. That recognition establishes research achievement. A buyer still needs a separate demonstration that the proposed system will improve its own operations.

The agent needs an examiner

One of Solvd’s more instructive AI projects concerns tax forms. At an unnamed payroll platform, analysts spent days mapping each new form’s fields to internal data. Solvd built an agent pipeline that groups fields, maps them to data-source templates, checks the results and retries failures. Humans review the uncertain remainder. Its evaluation compares the referenced data source against known answers, avoiding the mistake of treating matching wording as matching meaning.

The payroll case makes the testing heritage visible. The agent gets an examiner. The workflow has a route back to a person. Tracing includes cost tracking. For another business considering automation, the copyable move is to define correctness before celebrating speed. This approach needs accessible data, usable reference answers and someone qualified to judge exceptions; without those, an automated workflow can multiply ambiguity.

The economics require the same discipline. A reduction in testing costs describes one customer’s outcome, rather than the fee another customer should expect. A buyer evaluating custom engineering needs to count integration, review and maintenance alongside model usage. If an ordinary packaged tool already handles the task, commissioning a bespoke system may create more work than it removes.

Own the work around the model

Solvd sells business-to-business engineering engagements, supported by tools including Zebrunner, Carina and MCloud. Its current service menu groups work into Core AI, AI Data & Cloud, and Agentic Software Engineering. In market terms, the alternatives include an internal team, a specialist consultancy or a broader digital engineering provider. Solvd’s case rests on joining testing experience to research and implementation. Buyers should ask to see that combination in the proposed team.

“Companies should own their architecture rather than inherit it from whichever tool they bought first.”

Skylar Roebuck, CTO / September 2026

The line comes from Solvd’s OpenAI Select Partner announcement, which describes working inside clients’ codebases and agreeing success measures before handover. It is also a useful buying rule. Begin with one troublesome workflow. Specify the answer, the owner and the running cost. Then ask whether the system can keep doing the job when the demonstration audience has gone home.