Philipp Grobbel
18y · building since 14 · Stuttgart → SF (July – Sep)
ph.grobbel@icloud.com · +49 176 34645992 · linkedin.com/in/philipp-grobbel (never been active)
Started my first company at 14. What I've built since spans different markets, all connected by data science — a fairly classic arbitrage business that worked, then two venture bets that taught me why they didn't.
Currently in SF meeting people before starting Physics at TUM in October.
Projects
-
Resale operation — sneaker arbitrage → cross-platform pricing engine
Started at 14 with sneaker resale: in-store drops first, then automated checkout at scale — dozens of pairs per release. When demand for hyped products dropped, pivoted to the information layer: built a system that monitored retail inventory across selected verticals, aggregated prices across resale platforms via APIs, and modeled where spreads would open — mostly around special offers, product-generation sell-offs, and the occasional price error. Buying and selling decisions came from the model, not from watching drops.
What I learned: how markets work — aggregating the data that moves them, and understanding which factor matters when, across multiple sectors.
Wound it down in 2024: resale margins compressed as buy-side conditions on the marketplaces tightened, and operationally it had become a loop — the model was built, the rest was execution. Time was better spent on the next thing.
-
Autonomous selling agent for second-hand marketplaces
Chromium-based macOS app that sold your stuff for you on C2C marketplaces like Vinted, Kleinanzeigen and eBay. Photo in, listing out: generated description, upscaled image with cleaned background, a price forecast, and an agent that negotiated with buyers and handled shipping through the DHL API.
The interesting part was the data layer: a RAG aggregating price history, seasonality and weather effects, hype cycles, substitute products, next-generation releases and review signals per product — plus the condition of the specific item, and a seller profile built from listing history, writing style, response times, and further information about the seller. That fed the negotiation LLM its price corridor and what to emphasize. Planned second revenue stream: aggregated pricing signals for e-commerce companies to reprice returned goods.
Ran end-to-end, never publicly launched. We shut it down over two problems we couldn't solve: keeping marketplace sessions stable at scale — the agent had to refresh immediately to react — and the EU AI Act's transparency rule — an agent that discloses itself as AI violates Kleinanzeigen's terms of service. When it became clear the core mechanic couldn't be made compliant, we stopped.
Stack: Supabase (data + auth), Qdrant for the RAG's vector search, deployed on Vercel.
-
Matching engine for student jobs & internships
Job platforms treat students as keyword lists. I built a matching system that worked from actual profiles instead: students defined what mattered — specific experience to gain, income, commute, culture, or custom combinations — and the engine optimized placements against those weighted preferences, reusing the actor-modeling approach from the marketplace agent: understand both sides of the market, then personalize the match. Around the match itself: researching openings, a simplified application flow, interview prep, and advice on offers and contracts. The moat was meant to be data superiority over a plain LLM — in practice it never really got there.
Solo, shut down before launch — three walls this time. Modeling both market sides meant continuously aggregating profile and company data from many public sources — gray territory: using existing profiles and their career paths to generate recommendations, while technology keeps redrawing those paths. The matching complexity on top outgrew a solo build. And the unit economics never closed: HR-side CAC was far too high for realistic onboarding — companies won't pay to integrate yet another intern-sourcing channel. On top of that, a defensible moat against plain LLMs was getting harder to see.
To me, that's the point of ML and data science: clarity — physics becomes more graspable, markets become more efficient, and people can have a broader view for better decisions.
Off the laptop
Marathons and a 17:02 5K, road cycling, skiing.