-
-
Synthetic demo: review shared reporting gaps before replacing individual probes. A human decides the next action.
-
Synthetic replay: a persistent shift is evidence for human review, not proof of hardware damage.
Inspiration
A sensor alert does not tell a monitoring provider whether to inspect a gateway, replace a probe or charge a customer. My agricultural sensor research inspired FieldFix: a practical workspace for the decisions that happen after an alert.
What it does
FieldFix helps monitoring service providers review sensor evidence, approve a service quote, check recovery and record an approved settlement. The farm customer is the buyer; the monitoring provider supplies the replacement service. The demo quote is a fictional NZD 25.
Three synthetic cases show a shared reporting interruption, a persistent reading shift and a normal control. The operator reviews the evidence before selecting an action. The paid route then separates quote approval, buyer authorization, recovery checks, human acceptance and charging. An alert alone cannot buy anything. Insufficient recovery data blocks acceptance and charging.
The provider records acceptance in this single-operator prototype; it is not an independent customer sign-off. Warranty requests are drafts only, with no merchant adjudication or refund integration.
PayPal and AI
PayPal Sandbox Orders v2 and Payments v2 provide authorization, capture, verification and voiding of uncharged authorizations. The server checks identifiers, status, amount and currency. An uncertain response needs reconciliation, not an assumed success. PayPal is the settlement step in the service workflow, not simply a checkout button attached to an alert.
A local multiclass logistic-regression model, trained on 3,200 generated synthetic sequences, extracts 12 features from 72 hourly observations per device. It recognizes normal, persistent-shift, stuck and dropout patterns. Input checks, abstention rules and independent checks constrain the result. Scores are not calibrated fault probabilities: irrigation and a sensor offset can look alike. Versioned evidence ties each human decision to the observations and model used.
How I built it and what I learned
Python, SQLite, HTML, CSS and JavaScript power the application; NumPy is used only for offline model training. The shipped runtime needs only Python's standard library. OpenAI Codex assisted coding, tests, localization and documentation. Video narration is synthetic Niki speech generated through Runway with ElevenLabs Multilingual v2; it is a production tool, not the application's AI.
The challenge was preserving uncertainty without making the daily screen difficult to use. I learned to separate an anomaly, a human service decision, evidence of recovery and a payment outcome. Each can succeed or fail independently.
Progress during the hackathon
The case workflow, synthetic classifier integration, evidence-linked reviews, recovery gates, PayPal Sandbox integration and five-language interface/reports were developed during the event. v0.11 improves case progress labels and makes payment and acceptance roles explicit. The earlier academic research inspired the problem; its private LSTM, collaborators' models and research datasets are not included or used by this app.
Try it
Source: https://github.com/Yidizhu/fieldfix
With Python 3.10+, run python app/server.py --port 8767 --offline-demo and open http://127.0.0.1:8767/. No payment account or credentials are required for this local rehearsal. Use FF-102 and test the insufficient-data branch before the recovered sample. The README and reviewer quick-start guide explain a separate Sandbox run with your own test accounts.
The current integration completed a separately verified Sandbox authorization and capture on 11 October 2026 using virtual NZD 25 after simulated acceptance. The demonstration distinguishes local rehearsal from separately recorded Sandbox evidence; it does not claim continuous payment footage.
What is next
Validate terminology and acceptance/payment responsibilities with service teams, then evaluate on permissioned field data. All bundled readings, quotes and recovery checks are synthetic. There is no live sensor ingestion, real dispatch, production user isolation, real repair certification or customer pilot. English, Chinese, Spanish, French and Portuguese support does not establish global live-payment availability. The innovation is the reviewable handoff from uncertain evidence to approved service settlement; I do not claim unique fault detection or measured cost savings.
