BUILTPrototype
DoorLens
A local tool that classifies recorded train-door movements and tests whether a result holds up.
Snapshot
- My role
- Developer
- Timeframe
- September 2026 · hackathon build
- Status
- Prototype
- Organisation
- Nebula X 2026 hackathon
The problem
Nebula X 2026 hackathon submission (PS3). Classifies recorded door movements with a frozen model, re-runs a cycle under explicit synthetic recording changes, and ranks a train’s air-conditioning units by refrigerant-leak likelihood.
How the project ran
I owned
- Built the FastAPI backend that validates uploads, segments cycles and runs frozen-model inference
- Built the Next.js and Plotly front end, served as a static export by the same API
- Added “challenge this result” probes that copy and rescale a cycle and re-run the same model
- Wrote model-parity, API and browser verification scripts
Constraints
- The test data had no ground truth, so it was used only for format checks and inference, never for fitting
- Runs entirely locally: no database, cloud service, authentication or external AI API
Outcome
- 32 Python tests passed and frozen-model parity reproduced across 2,091 comparisons
- Sensitivity probes never alter the original prediction exports
Built with
- Python
- FastAPI
- NumPy
- pandas
- Next.js
- Plotly
- Playwright
Evidence
Related work
BUILTLive
ConnectSphere
Event networking with a sponsor-facing ROI report, not just a chat room.
- Role
- Developer
- Timeframe
- Personal project
BUILTArchived
EventureSG
A Telegram bot for finding local events.
- Role
- Developer
- Timeframe
- Personal project