# cashtime — case study by Anton Kozionov > Product owner and designer of a car-collateral loan app; I also built its support bot and a 28-city site network - Role: Product owner & product designer, the only PO on the team - Team: About six: CTO, back-end and front-end developers, QA; plus outside web devs. One backlog in ClickUp - Timeline: July 2023 – now - Tags: fintech, product owner - Case study page: https://kodsgn.com/work/cashtime/ ## Highlights - Designed the iOS & Android app (4.8★ App Store, 4.5★ Google Play) and the website - 28 city sites on one template; a new city takes one command - Built an AI support and application bot (not launched yet) cashtime lends money against a car, fully online, and the borrower keeps driving. Since 2023 I've written and prioritised the backlog for its developers, with the owner signing off on the big bets, and I design the app they build and the website. When the roadmap outgrew the team, I built the rest myself with AI agents: a support bot and a 28-city site network. ## A loan is a chain of outside services, so I design it as one journey The borrower proves who they are with BankID, Ukraine's bank-based digital ID. Then come documents, the car's data, a check, a contract in an e-signature service and a lien in the state register. Only then does the money arrive. Each link is a different service with its own failures and waits. I design what the borrower sees, waits for and can fix at every link, so it reads as one journey rather than a stack of screens. Borrowers are in a hurry and wary of lenders, and a lawyer signs off every public claim. Every rate and promise has legal weight, in a chat reply as much as on a landing page. ## I shape what gets built next ### Planned around channels we could build, because we couldn't outspend a larger competitor In April 2026 a larger competitor issued 2.4 times more loans a month and bought its traffic, while we lived on search. With the owner I scored the growth plan by impact, effort and time to result. The top three were channels we could build ourselves: an affiliate programme, a support bot and regional sites. All three got built; none has a result to show yet. ### Ranked problems by how long they'd gone unfixed, not only how often With agents, I traced 719 applications through 12.7k messages in the team's chat. My briefing for the owner ranked each problem by frequency and by age, with a proposed decision. Entering a spouse's details came up 68 times and had been broken for over a year. It got its own screen in the bot's application flow. ### Made the lawyer's review a step in the pipeline, not a bottleneck The website said "10 years of experience" in 35 places, for a company founded in 2020. Fixing each by hand would let it drift again. So facts like age limits became variables from one source; the age range alone replaced 159 hard-coded spots. Agents flagged each spot, I approved a before/after list, and a verifier agent checked the result. Anything legal went to the lawyer at the list stage, so her review became one step in the flow instead of a queue at the end. The verifier exists because a fix can be a new fiction. "100% security" once became "we don't share your data", while the privacy policy lists who receives it. ## One backlog for everyone who builds cashtime The CTO, back-end and front-end developers, QA and outside web developers all work from one backlog that I write and prioritise. Most of the design I do myself; I also mentored the team's graphic designer. A design isn't done at hand-off. It's done when the real flow works, so I check each build against edge cases, errors and small screens. ### Prototyped analytics on a mock API shaped like the real one, so the team could port it The admin panel had no analytics. Rather than wait for back-end time, I built a three-tab dashboard on a fake API with the same shape as the real one, then wrote it up as backlog items. Developers can swap the data source instead of starting over. [Image: Funnel tab of the analytics prototype on mock data] [Image: Marketing tab with applications by source and platform, mock data] *Prototype on mock data; no real clients.* ### Agent QA files bugs the way engineers want them Browser agents test whether a flow works, whether every state looks right and whether it matches the business rules. Each finding comes back with severity, steps to reproduce, expected versus actual, evidence and a suggested fix. ### Started the next product, cashtime GO, from a design system For cashtime GO, an unsecured micro-loan (April–June 2026), I began with a design system that lives in both Figma and React. An agent pipeline then built the landing pages from its components, so every page is made of parts I designed. It's pre-launch. ## What I built with agents I direct AI agents in Claude Code. I write the brief and the rules, review what comes back and decide what ships. ### A support bot that answers from the knowledge base and never guesses your balance A Telegram bot and web widget that answers from the company's knowledge base, takes applications and hands off to an operator. The whole knowledge base fits in what the model reads at once, so I skipped building search over a document database. The bot works out fees on a hypothetical loan. What you actually owe comes only from live account data, and it never invents a date. It attaches the regulator's required disclosure whenever terms come up. It stops selling to anyone in crisis, gambling or under pressure, and it strips personal data. It's built; it hasn't launched. ### 28 city sites share one template and one source of numbers One WordPress template runs 28 city sites, and adding a city takes one command. Loan terms come from one shared document, not typed into 28 sites. On the home page, the main content now appears in 2.4 s instead of 17.6 s (PageSpeed 53 → 94); self-hosting the fonts alone saved 5 seconds. The sites stay hidden from search engines until each city has its own content and the owner says go, so there's no traffic effect to report yet. [Image: Social preview for the regional sites: "Money secured by your car, without leaving your car"] *Social preview image for the regional network.* ## Where it is now - **4.8★ on the App Store** (57 reviews) and **4.5★ on Google Play** (40) for the app I designed, September 2026. - **All three top growth bets built.** The city sites wait on local content. - **646 copy fixes live** on the site, with the lawyer's sign-off where it mattered. - **The bot is built, not launched.** No usage numbers yet. ## What I'd do differently - **Launch the bot smaller and sooner.** It spent three and a half months in the build without a real customer. A read-only FAQ widget could have shown in weeks what people actually ask. - **Start with fewer cities.** My own plan said ten first. I built all 28, and they're waiting on unique content. - **Label mock numbers loudly.** The prototype estimates visits as applications × 6, so one funnel step reads over 100%. - **Bring the lawyer in before the copy.** One review up front beats hundreds of edits.