GetNinjas is the largest services marketplace in Latin America, connecting individuals and businesses with trusted professionals in more than 500 service categories.
I led the product and design work on the new AI-Native Ecommerce Platform, which created a more reliable journey for customers and a more profitable business for professionals and GetNinjas.
- Product Lead,
- a hybrid role combining product design and product management
- Me and two engineers
- Reporting to the CEO,
- and working with shareholders, the CFO, Principal PM, Head of Marketing and Head of Operations
- From February to July 2025
- (MVP) and from February to July 2026 (MLP)
- Platforms
- Mobile-first Web, App, WhatsApp Agent, and MCP
A lack of trust
For 14 years, GetNinjas ran on a simple model: customers posted service requests, professionals bought those leads, and negotiation happened off-platform.
Over time, this created lasting pain points:
- Customers said “I don’t trust the pro”, “I had a bad experience” or “No one responded to my request.”
- Professionals, meanwhile, paid for leads from clients who only wanted a quote or never replied.
In early 2025, a potential partnership with Redion (formerly Europ Assistance), Brazil’s second-largest insurance services provider, meant I was asked to cross-sell its services within the lead-gen flow.
But I realised the partnership could go further and tackle the business’s structural challenges…
What if…
we built
e-commerce
for services,
solving most of
our customers’
pain points?
Upfront fixed pricing, a service guarantee, platform-mediated advance payment, and scheduling at checkout.
To get stakeholder buy-in, I used AI to build a small, deliberately rough but fully working prototype in four hours. It proved the concept was technically feasible, simple to execute, and able to deliver value quickly within our existing acquisition-campaign journey.
Designing the funnel
Once we had buy-in, I focused the MVP on validating the value proposition through the simplest possible funnel. After six months, with solid product–market fit signals, I shifted to creating a frictionless, reliable and remarkable experience, preparing the product to scale.
MVPFocusing on product–market fit
Paid media Existing channel
Service page
Booking form
Gateway's UI Payment partner
Summary page
Provided by the partner
MLPFocusing on scaling the experience
Discovery Home, categories, search, agent, location by postcode or GPS, promo pages
Service details Scope, price, availability, customer reviews
Cart Multiple bookings, booking expiry, inline and social sign-in, saved address
Own checkout Pix and card, saved payment preference, designed to feel safe
Account Order tracking, support, profile
Tracking From the account
Rating CSAT, support
-
Find a service
MVP Paid media, existing channel
MLP Discovery. Home, categories, search, agent, location by postcode or GPS, promo pages
-
Check details
MVP Service page
MLP Service details. Scope, price, availability, customer reviews
-
Schedule
MVP Booking form
MLP Cart. Multiple bookings, booking expiry, inline and social sign-in, saved address
-
Checkout
MVP Gateway's UI, payment partner
MLP Own checkout. Pix and card, saved payment preference, designed to feel safe
-
Confirmation
MVP Summary page
MLP Account. Order tracking, support, profile
-
Service execution
MVP Provided by the partner
MLP Tracking. From the account
-
Post-service
MVP Provided by the partner
MLP Rating. CSAT, support
Designing trust
After six months running the MVP, we had real behaviour data and customers’ own words. Interest was high, but parts of the journey still left people confused or unsure.
Customers didn’t think in the categories providers used to sell services. So I designed a service structure around customers’ problems, and made it the foundation for search, navigation, browsing and the agent. Customers could describe what they needed in their own words and find services available at their address.
Collecting booking details was the biggest UI challenge: each service needed its own address, date, time, preparation and agreement, and the cart combined several in one payment. I turned every input into a chip that shows what you’ve picked and expands when you tap it, so customers only dealt with the detail they needed at each step.
Payment, which had been our biggest drop-off point, now happened entirely inside GetNinjas, with clear communication and fast access to previous preferences. Begin checkout to purchase conversion jumped from ~33% in the MVP to ~47% in the MLP, and average time on task fell from 322s to 131s (median 72s).
After payment, customers could come back to check the order’s status and scope, contact support and leave feedback.
The relationship of trust now had somewhere to live.
Designing with AI
I led the adoption of an AI-assisted way of working, which I built with the two engineers. We used spec-driven development inside a shared monorepo, where product specifications, executable prototypes, issues and review evidence lived alongside the product. A harness connected the design system, Storybook, decision records, specialised agents and automated checks.
The two engineers owned production architecture, implementation and runtime quality, while I owned product direction, interaction design and design acceptance.
The main challenge was how can we overcome AI slop and build brand recognition and quality perception through the interface, while still benefiting from the speed of generative AI?
After trying several approaches, I found that quick hand-drawn wireframes worked best for setting the agents’ direction, and behavioural specifications and high-fidelity code prototypes made it precise. The harness tied them together.
The StateDebugger made the prototype easier for people to inspect. Product, design, engineering and stakeholders could open any route and switch conditions such as reservation expiry, anonymous identity and offline availability, without replaying the whole journey.
Synthetic personas were a separate form of automated testing. Through a PostHog MCP connection, I brought behavioural data and session replays into the workflow, and aggregate data showed which behaviours recurred. I used that evidence to refine the personas and the tasks they attempted in the prototype.
The personas increased coverage but did not replace customer research or human review. The StateDebugger supported that review, while the repository kept decisions and evidence traceable even when features followed different paths. AI was part of how we built the product; the live customer journey did not rely on a model at runtime.
A new revenue stream,
with a stronger value chain
The rebuilt product brought discovery, booking, payment and account management together, with higher payment completion and a live, transaction-based revenue stream.
MVP Released in June 2025
Revenue per request above the lead-gen model.
Services completed in the first month.
Sean Ellis score.
MLP Released in July 2026
Higher session-to-order conversion.
Qualified leads captured, through inline login and cart recovery.
Net Promoter Score.
Periods and traffic vary; figures are averages from the first months.
Key takeaways
Sometimes people need to see an idea in action before they believe in it.
Fixed-price e-commerce was an old idea at GetNinjas, held back by doubts about feasibility and effort. A quick working prototype won buy-in by showing it could be done.
Product metrics alone don’t reveal design problems.
Acquisition sources, previous visits and customers’ wider path to purchase helped me interpret intent and identify usability issues. That gave us more specific behaviours to measure when evaluating design changes.
How a product responds shapes how customers judge its quality.
Prototyping in code let us test loading times, touch response and transitions during design. These details add up to how smooth and well made the product feels.
Shipping faster can outpace your ability to learn.
AI let us ship new features before we had results from earlier releases. That pace risked growing the product faster than we could learn from it and adjust direction.
Avoiding AI slop starts with the thinking you do before prompting.
I lost time correcting decisions I’d left to AI, even when I knew what I wanted. Choosing a direction first and specifying those decisions gave the AI a clearer brief, while leaving it room to suggest alternatives.