Building Lumosofy: Notes from InnovNation Serbia 2025

In December 2025, on a stage in Belgrade, Lumosofy — our autonomous AI sales agent for Shopify — took 2nd place at the InnovNation Competition Serbia 2025. This post is the story behind that result: why a consultancy decided to build a product, how a deliberately small team shipped it, and what standing in front of judges taught us that client work had not.

Why a consultancy builds a product#

Nilovate is an implementation company. We put AI agents to work inside other people's businesses, and our operating philosophy fits in one sentence: we build AI that delivers measurable business outcomes, not impressive demos.

That philosophy is easy to state and expensive to hold. Client work tests it constantly — but client work also has a limit: you inherit someone else's constraints, someone else's process, someone else's definition of done. At some point, if you claim to know how agents should be built, you owe yourself the harder test: build one end to end, own every decision, and let the market grade it.

We chose e-commerce because it is where the gap between "AI feature" and "AI outcome" is most visible. Every Shopify store has a chat widget. Very few of those widgets sell anything. Shoppers arrive with real questions — will this fit, how does it compare, when will it arrive — and when nobody answers at 2am, they leave. The store did the hard work of getting the visitor; the last meter failed. That last meter felt like an agent-shaped problem.

What we built#

Lumosofy is an autonomous AI sales agent for Shopify. Not a chatbot with a product feed bolted on — an agent whose job is the sale:

• Native Shopify integration with one-click install — no development project to get running
• Automatic catalog sync, so the agent's product knowledge is always current, not a stale export
• Conversation memory, so a returning shopper is a continued conversation, not a stranger
• Agentic checkout — the agent can move a shopper from question to completed purchase, not just answer and hope
• 24/7 availability, because the 2am shopper is exactly the one nobody else is serving

The brand line we settled on — Intelligence. Refined. — was a design constraint as much as a tagline. Every time a feature idea felt clever but unhelpful to the actual sale, it lost the argument.

How we built it#

Nilovate does not have an office building or a headcount chart. We run as a lean, borderless network of specialists across 15+ countries — remote-first, assembled per problem. Lumosofy was built the same way we build for clients, with the same two disciplines doing most of the work:

Weekly demos, no exceptions. Every week, working software in front of critical eyes. Not slides — the agent, live, with whatever was embarrassing still embarrassing. Weekly demos are how a distributed team stays honest: you cannot hide behind a status report when the thing has to run on Friday.

Production-ready from the first commit. The stack is the one we trust in client work — Next.js, TypeScript, Claude, Shopify, Vercel, Supabase — and the standard was "a merchant could run this," not "this will survive a demo." That choice looks slow in week two and fast in month four, when pilot merchants show up and there is nothing to rewrite.

By the time the competition appeared on our radar, Lumosofy was not a concept. It was running in pilot deployments with real stores, real shoppers, and real numbers accumulating.

The competition#

InnovNation Serbia 2025 gathered some genuinely impressive teams, and pitch competitions reward a specific skill — compressing months of work into minutes of attention — that has surprisingly little overlap with building.

We made one strategic decision early: lead with the numbers, not the vision. Our pilot deployments were showing a +32% average conversion uplift — pilot data, from pilot stores, and we said so on stage rather than rounding it up into a universal promise. The vision slide came last, almost as a footnote.

That choice cut both ways. Vision-first pitches soar higher in the room; numbers-first pitches survive questions better. The judges pushed exactly where they should have — how does this scale beyond pilots, what happens when the agent is wrong, why won't the platform build this — and having lived answers instead of prepared ones is what carried us.

Second place. A better team-building exercise than any offsite we could have designed, and a result we are genuinely proud of — with the useful sting of not being first keeping us honest.

What we learned#

Lessons we took back into client work, in rough order of importance:

1. Outcomes are a moat; demos are not. Half the field had impressive demos. Pilot numbers were rarer, and the difference was visible in the Q&A, not the presentation.
2. Honest framing outperforms inflated framing. Footnoting "+32%" as pilot data cost us nothing and bought credibility for every claim after it.
3. Small and senior beats large and coordinated — for products at this stage. A network of specialists across time zones shipped faster than a co-located team twice the size would have, because nobody was waiting on anybody.
4. Constraints are a feature. No office, no big budget, weekly public accountability. Each one felt like a limitation and functioned as a filter against waste.
5. Eating your own cooking changes the recipe. We advise clients differently now — with the specific humility of people who have shipped an agent into production and watched it meet real users.

Early access is open#

Lumosofy came home from Belgrade into the next phase: early access is now open for Shopify merchants. If you run a store and want an agent that treats the 2am shopper as seriously as the 2pm one, the Lumosofy page has the details and the early-access list.

---

And if your interest is broader than Shopify — if you are wondering what an agent built with this philosophy would look like inside your own company — book a 30-minute intro call. We will bring the lessons; you bring the process that annoys you most.