Four systems, one open-source project, and the decisions behind them.

A balance check can say a purchase is affordable while a bill due before payday still pushes the account below its minimum. Trevad forecasts every day of the next 90 and recommends the safest way to pay.

• Day-by-day 90-day cash-flow forecast rebuilt from transactions and messages
• Recommends paying in full, in part, by seller installments, waiting, or not proceeding
• Every candidate plan is replayed and only plans that never breach the minimum balance can be chosen
• Messages can change cash flows, never the decision; instruction-like text is detected and ignored
• Deterministic decisions checked by a separate validator before they are returned
• CLI, FastAPI API and web interface all call the same decision engine
• pytest scenario and product tests

Payment operations need visibility into provider attempts, latency, status, and who is authorised to act on a payment.

• Angular full-stack operations workspace
• Flask API and MongoDB
• JWT authentication
• Merchant, finance, and admin roles
• Server-side merchant data scoping
• Provider-attempt history
• Search, filtering, sorting, pagination, dashboards, and analytics
• Automated frontend and backend tests

Field researchers need to submit structured environmental observations with photo evidence into a searchable cloud system.

• Express API deployed on Azure App Service
• Cosmos DB for metadata
• Blob Storage for images
• API-key protected routes
• Timing-safe API-key comparison
• Application Insights telemetry
• GitHub Actions deployment
• Deliberate cost and scaling trade-offs

Tamper detection often sends raw sensor data to a remote server, increasing latency and privacy exposure. SecureSense performs inference on-device.

• Arduino Nano 33 BLE Sense
• Motion, light, and proximity data
• 45 deterministic IMU and light/proximity features
• 100-tree random forest running on the device
• Independent session split to reduce leakage
• Held-out Session 3 result: 83.33% accuracy and 0.8324 macro-F1
• Five-window live voting result: 86.67%
• Mean on-device inference: 236.68 µs
• Local Flask security console for evidence logging