Management student (MIM @ NOVA SBE, Lisbon) who builds and ships software — mobile apps, internal business tools and AI agents, from first sketch to production.
- ImmoApp — property-management platform for private landlords and property managers. iOS app + API with tenant, owner and manager roles, rent tracking and tax reports. In beta on TestFlight.
- Facility-service tooling — internal web apps for a facility-services company: quote calculator, fleet management with fuel-card fraud checks, and a client portal. Running in production.
- Apto — built at the START Unicorn Lisbon hackathon: students solve real work samples from startups instead of writing cover letters.
- FITFIT — iOS app built on the iPhone's LiDAR scanner.
- Ironman Trainer — personal AI coach that syncs my Garmin data and plans triathlon training.
Most of this lives in private repos — happy to walk through the code on request.
- footnote — question answering over 10-K filings with citations, and an eval harness built from XBRL data that measures how often it's actually right.
- honest-odds — are prediction markets calibrated? Calibration curves, favourite-longshot bias and a strict out-of-sample check on 2,946 resolved Polymarket markets.
- touchdown — vision-based drone landing: detects the pad, estimates pose with PnP and a Kalman filter, and guides the descent. 86/100 simulated landings, median error 9.9 cm, real-time on a laptop CPU.
- hivemap — aerial crowd analysis from drone video: person detection, a hand-built multi-object tracker, density maps and ground-plane analytics, measured on the VisDrone benchmark.
- emberwatch — fire and smoke detection with temporal alerting, evaluated on false alarms as much as on recall. A portfolio project, not a certified safety device.
- spike-and-gradient — linear regression, a Hodgkin-Huxley neuron and a neural network (97.8% on MNIST) from first principles in NumPy, each verified against a trusted reference.
- blind-committee — a multi-agent LLM "hedge fund" (analysts, risk manager, portfolio manager) that anonymises companies so the model can't remember what happened. Honest result: it doesn't beat equal weight.
- alpha-graveyard — where discovered trading strategies go to die: 15,000 candidates from a genetic search, zero survive the held-out decade once the overfitting statistics from fluke are applied.
- fearcurve — recovers the market-implied distribution of an asset from option prices (Breeden-Litzenberger), validated against the analytic Black-Scholes case.
- fluke — is your best backtest real or just the luckiest of N tries? Probability of Backtest Overfitting (CSCV) and the Deflated Sharpe Ratio after López de Prado.