Francisco Rau_

Citania — booking & payments for LATAM beauty

The product the research pointed to: a multi-tenant booking-and-payments platform in the Fresha category, built for Latin America (Spanish, MXN, CDMX-first) — a genuine pain point with no serious player in the region yet. A TypeScript monorepo — Express 5 + Postgres, contract-first OpenAPI, a React web app and an Expo React Native mobile app.

• 100+ integration tests over the full booking flow, with database-level slot exclusivity — double-booking is impossible by construction, not by hope.
• Two-way Google Calendar sync live, with loop prevention via source tags, booking IDs and etags.
• Role-based pricing, cancellations, reviews, reminders, segmented marketing + loyalty, and a full owner dashboard (calendar, services, team, inventory, clients, reports).

Bookings · day calendar⤢ Team shifts · scheduler⤢ Reports · revenue & KPIs⤢ Marketing · segmented email⤢ Team · roles & pricing⤢

Weatherbot — mispricing detection on weather markets

Weather markets settle on official observations published on a fixed schedule, so in principle mispricings should be detectable. I built the system end-to-end to test it: forecast ingestion, implied-probability extraction from order books, no-arbitrage / coherence checks across every temperature strike ladder, and latency measurement between an observation publishing and the market repricing — backtested across ~3,000 markets.

The finding is negative — and that is the point. The original thesis went 0 for 24 against resolution, and the book spread on those contracts runs 30 to 100 percent of mid, so nothing was tradable even where an edge looked real. Exchange fees are zero on that venue; spread is the cost. I documented it as a rigorous null result instead of curve-fitting a backtest that "works." Knowing when a signal isn't tradable is the skill — it's the same discipline that decides whether a metric is trustworthy enough to put on a dashboard.

Alpha Scout — find undervalued players before the market does

A hedonic fair-value model (R² = 0.74 on live data) prices every player from performance, age, league difficulty and contract; calibrated probabilities say how often profiles like his actually rise. Forward-tested, not curve-fit: players the model flagged as undervalued were signed 1.7× more often than their peers in the following transfer window.

• The data is the moat — automated match-by-match harvesting across 155 leagues, self-healing browser pipeline, weekly refresh, ~98,000 player-seasons of history with observed 12-month outcomes.
• Honesty as a feature — the backtest is published per league including where the model is weak, and it declines to price players above €60M: outside its validated range, it says so instead of guessing.
• Built for the buyer — club view with squad gaps and realistic budgets, "similar players under €1M from your region", contract-expiry radar, one-click club reports.

Explorer · scout ranking & probabilities⤢ Player profile · fair value & form⤢ Club view · Boca Juniors sample⤢