João Vaz
joaosfvaz@gmail.com · linkedin.com/in/joaovaz00 · joao-vaz.com
Product Manager, OLX Group
I build tools for 12,000+ car dealers and 10M+ visitors.
Tools for car dealers by day. After hours, chess, sport and a camera.
Product Manager at OLX Group (Motors, B2B), part of Prosus. I set the strategy for a 10-person team and build my own agents.
- +150%
- weekly engagement across 3 data products, in 18 months
- 12k+
- professional car dealers in Portugal, Poland and Romania
- 40+
- experiments: A/B, AA and fake-door
- 75+
- in-person dealer interviews
- 90+
- PMs trained on AI tooling
What I do when the laptop closes.
I build tools for 12,000+ car dealers.
I am a Product Manager at OLX Group (Motors, B2B), part of Prosus. My team builds the products dealers use to price and sell their cars, and I lead the AI agents for our domain in the dealers' assistant.
So far: +150% weekly engagement across 3 data products, in 18 months; 75+ in-person dealer interviews; and 90+ PMs trained on AI tooling.
How I work
- I go and see the dealer.
- More than 75 interviews in three countries. A solution meets dealers before it goes into delivery.
- I split the problem before I pick a solution.
- By the job the user is trying to do, with more than one end state on the table.
- I write short, and the conclusion comes first.
- Each option sized, each choice with its reason, and tickets that anyone can test.
- I measure outcomes, not output.
- A key result says what changed for the dealer, not what we shipped.
- I build the tools I need.
- Agents, Claude skills and the side projects further down.
Six products, one job.
Help a dealer buy the right cars, price them well and sell them faster. I own the strategy across all six, from the data pipelines to the screens.
I lead my domain's agents in the dealers' AI assistant. They answer questions about prices, performance and the market, from our own data.
- Agent specs
- MCP tools
- Evals
- Traces
What each car is worth and how fast cars like it sell, next to the dealer's asking price.
- ML model
- Price-to-market
- Time to sell
Supply, demand and prices for every segment of the market, down to the competing listings, so a dealer knows what to buy next.
- Supply & demand
- Days of supply
- Competitor listings
Views to calls for every car, benchmarked against similar cars. Every dealer gets it, on every plan.
- Funnel metrics
- Benchmarks
How the dealership is doing at a glance, with alerts that lead straight to the cars that need action.
- Stock health
- Alerts
- Multi-branch
One next step per car, from the signals of every product above: renew it, reprice it or give it more visibility.
- Pricing
- Renewal
- Visibility
An agent change is judged on evidence, not on a demo.
dealer“Why isn't my Audi A4 getting leads?”
- route → pricing & performance agent38 ms
- tool · listing_performance(car)212 ms
- tool · fair_price(car)187 ms
- mandate · read-only ✓2 ms
- answer · grounded in 2 tool results1.4 s
assistant“It's priced 9% above similar A4s and views fell after day 12. Cars like it sell in about 24 days at market price.”
- correctness
- groundedness
- scope
- plain language
- Evals
- Test questions with known answers for my domain, graded on correctness, groundedness, scope and refusals, so a change is judged on evidence, not on a demo.
- Traces
- I read agent runs span by span to catch invented values, routing errors and the same question getting two different answers.
- Mandates
- What an agent may do for a user: read-only or acting, whether an action can be reverted, and when a human must confirm.
- Automation by risk
- Each capability gets the lightest level of automation its risk allows.
- Tools & MCP
- Domain data behind MCP tools the agent calls, so answers come from the source, not from the model's memory.
- Context
- Personas, playbooks under a token budget, what an agent is not for and where it hands over.
- Human-in-the-loop
- Human review on critical paths until there are enough traces to trust an LLM judge.
- Skills for PMs
- I wrote the Claude skills rolled out to every PM in our unit, to define their domain's agent and to decide when a change needs a spec.
Helping dealers sell cars, since 2023.
Strategy for dealer insights, pricing and recommendations, from the data pipelines to the screens, with stakeholders in five markets.
Outcomes
- +150%
- weekly engagement, 3 data products
- ~72%
- average OKR completion, 2.5 years
- 40+
- experiments: A/B, AA, fake-door
- 90+
- PMs trained on AI tools, 75% now use them
Deliveries
- AI agents for my domainDefined what my domain's agents in the dealers' AI assistant can answer about pricing and listing performance, and tested them with evals and trace reviews.
- Market insights redesignRebuilt the view where dealers see supply, demand and prices for every car segment, on desktop and mobile.
- Segment drill-downLet dealers go from a market segment straight to the competing listings in it, with price, mileage and days on market.
- Consistent listing statisticsMoved listing statistics to one data source, so leads count the same in the product, in sales reports and in BI.
- Product workshop for ~100 peopleCo-designed a two-day workshop for 8 teams, with 8 Claude skills that turn dealer interviews into a scored roadmap.
- AI skills for every PMWrote Claude skills that help each PM define what their domain's AI agent should do, and decide when a change needs a spec.
Same team and scope, on the machine-learning side of the product.
Outcomes
- 75+
- in-person dealer interviews in PL, RO and PT
- 12 mo
- to promotion to Product Manager
Deliveries
- Fair-price estimatesBrought a machine-learning price estimate into the product, so dealers see what each car is worth next to their asking price.
- Time-to-sell estimatesAdded how long similar cars take to sell, estimated from a car's specs.
- Dealer researchInterviewed dealers in person in Poland, Romania and Portugal to understand how they price and sell their cars.
Outcomes
- 6 mo
- from trainee to Associate PM
Deliveries
- Lead-generation discoveryResearched how professional dealers get their leads, to find where the product could help them get more.
- Recommendation experimentsRan A/B tests with Data Science on which types of recommendation to show next to paid-visibility products.
A month-long AI programme by Prosus and the THNK School of Creative Leadership, with people from across the group.
Outcomes
- Selected
- from companies across the Prosus group
- 1 mo
- full time in Amsterdam
Deliveries
- Seller-inbox prototypeBuilt with my team a prototype inbox where an AI assistant screens out buyers who aren't serious.
- Live demo on stagePresented the prototype on stage to the whole programme in Amsterdam.
Toolkit
Agentic AI
AI-assisted building
Data & experimentation
Delivery & technical
CRM & go-to-market
- Education
- MSc Industrial Engineering & ManagementUniversity of Minho & TU Graz, 2021 to 2023
- Languages
- Portuguesenative
- EnglishC2, proficient
Things I built myself.
Tabiya
A chess coach that downloads my games, analyses them with Stockfish and explains every move in plain English.
Read more →
PM-OS
My team's knowledge base, kept true by seven scheduled Claude routines.
Read more →Team-context MCP
A read-only MCP server, so the company's AI assistant can answer questions about my team's work.
Read more →Portugal Explicado
Portugal's taxes, laws and economy in plain Portuguese, every number linked to its official source.
Visit the site ↗Makes sense? Email me.
obrigado, João
Quick answers
- Current role
- Product Manager, OLX Group (Prosus)
- Experience
- 4 years in product
- Languages
- Portuguese (native), English (C2)
- Based in
- Lisbon, Portugal
- Work setup
- Hybrid
- Relocation
- Open to relocation
- Right to work
- Portuguese and EU citizen: no visa in the EU
- Notice period
- 30 days
- Open to
- Product Manager, AI Product Manager, Data Product Manager and Product Builder roles
Off the clock?
- Level
- XP
- Achievements
- Photos
- Easter eggs
- Time