A Flutter/Android transit-alarm codebase, with its development history, tests and underground-tracking experiments, prepared for AI training and software-engineering evaluation.
I built GeoWake so I could sleep, read or switch off on a journey without missing my stop. Choose a destination and how early to be alerted; the app tracks the route and prepares the wake alarm. The code covers pre-boarding, distance/time/stops alarms, deviation and rerouting, background recovery, native Android delivery and two Node backends.
Underground tracking became the hardest part. My initial aim was to combine GPS and phone motion to follow the whole route, continuing on IMU measurements when GPS disappeared. Ordinary movements of a phone in a passenger's hand or pocket made that unreliable. The work progressed from raw acceleration integration to a route-constrained EKF, then toward detecting the train's braking, dwell and departure patterns and connecting them to known stops.
The archive includes that applied research: estimation and uncertainty code, ZUPT and station association, conservative alarm bounds, replay tests, diagnostic reports, failed approaches and historical fixes. The stop-event approach remained experimental. I stopped pursuing the product because obtaining enough suitable data to tune and validate it became too expensive. The research history separates implemented code, reported offline findings and work that remained open.
I am offering the existing code and engineering records for model training and software-engineering evaluation. No ongoing product-development engagement is included.
Data scope: mixed-source metro catalogs and certain route/scale data are withheld. Labeled schema stubs preserve code interfaces. Read the export scope and verification effects.
| Review question | Guide |
|---|---|
| What is offered and why is it useful? | Buyer brief |
| Where is everything? | Buyer index, feature status |
| What happens from search to wake-up? | Journey, pre-boarding, deviation and rerouting |
| Why did the underground approach change? | Underground evolution |
| What about purchases, sharing, guardian, anti-theft and data? | Supporting systems |
| How do inputs become state, storage and outputs? | Data-flow and schema map |
| Which decisions are documented or still uncertain? | Decision register |
| How can I run the checks? | Reproduction notes |
| Which claims have evidence? | Claims register |
| Where are the measured coverage reports? | Coverage and scope, LCOV, Cobertura |
| What history and local fixes are included? | Branch history, local capture |
| What is excluded or separately licensed? | Rights and scope |
Watch the historical demo or run the two historical exercises.
The delivered evidence bundle supplies exact-revision file hashes, a source and test navigation index, raw check receipts and two verified historical fail-to-pass examples. The short narrated demo is a historical development simulation; it is not a field-accuracy measurement or proof of the final commit.
The verified available-code profile passed 1,423 Flutter, 61 Express and 16
sharing tests at 0752b78. Measured line coverage was 65.08% for Flutter
(10,362 / 15,923 lines across 133 instrumented files) and 72.66% for Express
(303 / 417 across 16). Eight dataset-dependent test files are excluded;
these are scoped measurements, not whole-repository coverage. Each subsequent
GitHub CI run
publishes its own exact-revision coverage artifact and summary.
Historical product, architecture and research documents remain traceable through the archive index. Their old prices, launch plans and performance aspirations are historical context. New explanations and packaging changes are dated September 2026; they are not reconstructed old tickets.
Raw recordings from friends' phones are a separate optional data candidate. They are not silently included in the code license. No users, revenue, service accounts, signing keys or production credentials are offered with this corpus.
python3 tools/sale/verify.py --profile code --out /tmp/geowake-verificationUse Flutter 3.44.6, Node 24.x and Python 3. The code profile explicitly excludes eight dataset-dependent files, including gates that would otherwise pass on empty data. Their original source remains present. See the reproduction and export-scope guides for the separate complete pre-exclusion source receipt.