Products used by 11,000+ people · measurable revenue impact · AI at scale

Szymon
Wesołowski

Seven years inside Poland's largest retail network. The last three spent moving AI out of the slide deck and into 13,000 stores.

A replenishment engine measured at about 1% incremental turnover. A data product past PLN 100+M. An assistant that now answers most of the network's operational questions without a human ever picking up. Before all of it, six years running a fifteen-person software house with my own P&L.

I also co-authored the company's EU AI Act governance framework – shipping AI at this scale is only half the job.

11K
franchisees served daily by products I own
~1%
incremental turnover, proven against 2,000 control stores
100+M
PLN from a data product built from zero in five years
70+
AI systems under the governance framework I co-designed
Flagship work
01

AI replenishment for 13,000 stores

Owned end-to-end with five engineers and three data scientists. Held the line on a six-month controlled rollout instead of scaling on early adoption signals.

Validated impact against 2,000 matched control stores before wider deployment, demonstrating approximately 1% incremental turnover. Custom LightGBM model on Azure.

02

Enterprise RAG assistant, 65% → 92% accuracy

Launch and optimization for 1,600 employees across 9,000+ internal documents. 30–50% monthly actives within six months.

Azure AI Search with OpenAI and Gemini models. Accuracy raised on a 100-question golden set by fixing input document quality and rethinking chunking. Self-reported saving of two hours per employee per month.

Rejected personal and HR data despite user demand – access-control and integration cost outweighed the benefit. Scoped the assistant to non-sensitive content instead.

03

Franchisee assistant, 150K+ queries

Now the primary information channel on store operations for 11,000 users – operational questions answered without human contact.

Balanced pressure for a faster rollout against the risk of unreliable answers by introducing golden-set evaluation before scaling – measurable answer quality, not stakeholder perception, as the basis for product decisions.

04

Retail data, zero to PLN 100+M

Cumulative revenue over five years with 60 FMCG partners.

Defined the commercial model, the pricing and the data-privacy boundaries together with Legal and the Commercial Department.

05In design · Q4 2026

Real-time voice agent for 11,000 franchisees

Problem framing through scenario-based journey design, conversation scope, escalation paths and the guardrail model.

Changed the design after discovery with franchisees showed full automation was not the right target for every scenario. Reframed the product around selective self-service, clear conversation boundaries and escalation to a human.

Also delivered – a marketing content generation suite (video, copy, comms) · document intelligence over 60,000 store lease contracts · a customer-feedback insight tool processing 300K+ items a month · an HR tool for individual development plans, 500+ users.

How I operate

Evidence before scale.

Golden sets, matched control groups, 100+ interviews with store staff and franchisees. Measured answer quality beats stakeholder conviction every time.

Governance as product work.

Co-designed the enterprise framework covering 70+ AI systems: EU AI Act risk classification and mandatory pre-deployment checks for data lineage, evaluation and human oversight.

Three teams, one AI Product Manager.

15+ engineers and 3 PMs today. Earlier, an enterprise rollout across 10 countries and 25 stakeholder groups – and a P&L of my own for six years.

Business outcome over feature output.

Scope is judged by what it moves – incremental turnover measured against 2,000 control stores, a data product built from zero with 60 FMCG partners, and features dropped where the integration cost outweighed the benefit.

Craft
Product

Strategy · discovery and hypothesis validation · roadmapping · OKRs · RICE · experimentation and A/B testing · funnel and cohort analysis · MVP to scale · P&L ownership

AI

LLM applications · RAG – chunking, retrieval, golden sets · real-time voice agents · document intelligence · evaluation · guardrails and human-in-the-loop · inference cost and latency · PoC to production

Stack

Azure OpenAI · Google Vertex AI · OpenAI, Claude, Gemini, Mistral · Langfuse · GA · Mixpanel · Pendo · PostHog

Governance

EU AI Act risk classification · AI governance frameworks · GDPR · vendor and build-vs-buy assessment

Before Żabka
2018–19
E-commerce Manager · LaboPrint SA

Four online stores launched to PLN 200k+ turnover in four months. A five-person team across development, content and performance marketing; reporting rebuilt around CAC and LTV.

2015–18
IT Delivery Manager · GlaxoSmithKline

A corporate portal delivered across 10 countries in LATAM and EMEA for 10K+ users, coordinating 25 local stakeholder groups and the regulatory review path in a validated pharma environment.

2014–15
Brand Manager · Internetowykantor.pl

Brand strategy built and executed; a loyalty programme that cut client churn by 8%.

2008–14
Chief Operating Officer · eFRESH

A 15-person software business with full P&L ownership – PLN 5M in incremental group revenue, 15 long-term B2B contracts including Pearson, Eurocash, PBG SA, DGA and Nestlé, and 100+ digital brand programmes. "Poznań Entrepreneurship Leader"; two-time winner of "Golden Issuer Website".

1999–2007
Project Manager · Usability & Marketing

Fresh Group, Symetria and others. An interactive agency team of five, key client relationships, usability audits and market analyses.

Beyond the role
Teaching & mentoring

50+ talks and trainings, including generative AI at internal events and the Puls Biznesu conference. Lecturer at the School of Banking and elsewhere, 2010–2018 – 40+ sessions. Mentor to 10 product managers; ADPL Certified Mentor.

Education & certification

University of Economics, Poznań – Economic Journalism & PR, 1997–2001. School of Banking – Trainers Academy, 2010–2011. PSPO II · Product Analytics and Product-Led, Pendo.io · ADPL Certified Mentor.

After hours

Aikido instructor – children, teenagers, adults. Freediving instructor. Both teach the same thing product work does: pace yourself, and know when to stop pushing.

Say hello.