In June 2026, I’m leaving for a 3-year trip. The rough itinerary: Philadelphia → a summer camp in the USA (lifeguarding) → Seattle → Fiji → Australia on a Working Holiday Visa → New Zealand → SE Asia backpacking → Canada. Somewhere in there, I finish a data science project that’s been running on a Pi in my bedroom since 2025.
That project is TravelNet.
What is it?
TravelNet is a live data collection, analysis, and visualisation system built around the trip. It runs on a Raspberry Pi (tucked away at home, not in my bag) and continuously ingests:
- Location data — GPS traces from iOS Shortcuts every 5 minutes, with Overland running as a continuous secondary tracker
- Health data — Apple Health exports: steps, heart rate, sleep, workouts
- Financial data — Revolut and Wise CSVs uploaded automatically, plus a cash endpoint for physical transactions
Everything goes into SQLite via a FastAPI backend, backed up to Cloudflare R2 with age encryption. Remote access is over Tailscale. The whole thing runs in Docker and has been through a dry run in Ireland in March 2026.
Why bother?
Two reasons, one practical and one honest.
The practical reason: I’m going to spend three years moving through a wildly varied set of environments — tropical cities, remote islands, mountain hostels, beach towns. That movement will leave traces: in my spending patterns, in my health metrics, in the density and nature of my GPS data. That’s a dataset worth having. No one else will ever have quite this dataset, because no one else will have quite this trip.
The honest reason: I want to get good at the hard parts of data science. Not benchmark datasets. Not Kaggle. Real data with missing values, sensor dropouts, timezone hell, multi-currency FX complications, iOS background execution bugs, and all the other ways the world resists being measured neatly. TravelNet is a machine for generating hard problems.
What will the ML look like?
I won’t start the analysis until I’m in Australia — I want at least three months of US data as a baseline before I begin. But the plan is:
- GPS segmentation — clustering and HMM-based approaches to identify distinct travel behaviours (slow exploration vs transit, urban vs rural movement patterns)
- Spend analysis — what does daily cost of living actually look like across different countries and travel styles?
- Anomaly detection — unusual spending events, atypical movement patterns, health metric outliers
- Time series regression — predicting future spend or movement based on past patterns
The dashboard is planned too: an interactive site with the full dataset visualised, plus a public anonymised version on GitHub.
Where it is now
The system is built. The Ireland dry run (March 2026) confirmed:
- Irish coordinates showing up correctly in location data ✅
- Dual-source GPS coverage working as designed ✅
- Manual cash transactions logged correctly ✅
- Cloudflare backup running and restorable ✅
The remaining pre-departure work is hardening: persistent cron logging, testing mid-upload internet drops, confirming the Pi auto-restarts cleanly after a power cut.