Sellers were signing up for Jumba Connect and then stopping. Of more than 50 signups, 3 finished a profile.
Jumba is a construction materials marketplace in Kenya. Buyers order through the Jumba app, and sellers list their products through a seller portal called Jumba Connect.
Without sellers listing stock, buyers had little to order. Getting sellers from sign-up to selling was the problem I was brought in to look at.
I traced the onboarding funnel one step at a time. More than 50 sellers signed up. 15 went on to create an organisation. 3 completed a profile. Roughly 9 in 10 sellers were lost before they could list anything.
Listing was the biggest hurdle. Adding one product meant typing 40 to 60 fields by hand. For a seller with a full stock list, that is a long stretch of data entry before a seller sees a single order. Sellers just want to sell, and Jumba Connect was one more sales channel for them.
I moved the sprint's goal from completed profiles to a seller's first sale. Profile completion only showed that sellers filled in forms. A first sale would show Jumba Connect was working for them. With that as the goal, getting products listed quickly mattered more than a complete profile.
I then rebuilt onboarding around listing. The journey became: verify your phone, list products, see your shop, then finish setup. Profile, payout and KYC details moved later, after a seller had seen what their shop would look like.
Listing itself became search-first. Most products sellers carry are not new, and Jumba already had a catalogue. A seller searches it, picks the product and adds only their own details, such as price and stock. They create a new product only when it is really missing, which also keeps duplicates out of the catalogue.
Sellers came in two shapes. A hardware shop might list 10 to 20 products by hand. A wholesaler works with hundreds of items, often in spreadsheets. I designed a guided path for the first and a bulk path for the second. The team's default for bulk was CSV upload. I pushed for an in-app table that works like a spreadsheet instead, because downloading, editing and re-uploading a file is extra work for the seller. Uploaded products are matched against the catalogue and sorted into matched, likely matched and unmatched, so obvious matches go straight through and the rest get a seller or Ops review.
For suppliers with more than one branch, price and stock are set per location, with defaults a seller can override.
On the buyer side, I designed and built an AI-assisted OCR shopping feature in React. A buyer scans a materials list and the app builds a cart from it.
The team's reaction to the first version was "okay, meh". It showed off the technology but did not make shopping easier. I changed the approach: every AI result became a suggestion the buyer confirms, and the flow was designed to cut the decisions in an order from more than 50 to around 10 to 15, using verification steps and progressive disclosure.
I handed it over as a demo-ready prototype, with a stakeholder presentation, a user testing guide and developer documentation.
I represented design in sessions with the product, engineering, operations and sales leads, on search, seller onboarding, payments and the AI cart. I also mentored a junior designer through weekly design reviews, asked for several options before settling on one, and gave structured feedback on their work to the founder.
Measure first sale, not profile completion. It kept the sprint on what sellers came to Jumba Connect for.
List first, paperwork later. Profile, payout and KYC details wait until a seller has seen their shop.
Search the catalogue before creating anything. Sellers add only what is theirs, and the catalogue stays clean.
Two listing paths. Guided listing for small shops, and a bulk table with catalogue matching for wholesalers.
AI suggests, the buyer confirms. A misread item on a materials list goes straight into someone's order, so every scan result is checked before it reaches the cart.
Design for low-end and worn phones. Most users work on the go or on construction sites, on low-end or worn smartphones.
The engagement ended in January 2026. Jumba had a search-first listing flow designed around first sale, and a working AI shopping prototype ready for development and user testing.
Putting my demo in front of people early taught me quickly why my first solution would not work. It took my own bias out of the decision and opened the discussion to solutions that mattered more to the user.
On Jumba Connect, the lesson was to cut every requirement standing between a seller and their first sale.
Testing an AI growth coach with small business owners, and finding three launch blockers before any customer did.
View case study 2024An internal tool had the right data but the wrong structure. I redesigned it around how operations teams actually work.
View case study 2022Introducing bank transfers as a payment method for Nigerian merchants who didn't trust card payments.
View case study