Porchlight

01 / Context

The problem

A detector output is only the beginning of an edge-vision product. Repeated detections can overwhelm an operator, and a temporary network outage should not silently erase the event history.

02 / Ownership

My contribution

I prototyped a C++17 runtime with temporal aggregation and alert suppression, a SQLite offline queue, and a FastAPI/backend-dashboard path. A replay harness makes the event flow inspectable before a live-camera deployment.

03 / Architecture

How it works

1. 01

   Ingest

   Read replayed or synthetic detections into the edge runtime.

2. 02

   Aggregate

   Group related events and suppress duplicate alerts over time.

3. 03

   Queue

   Keep pending event information in a local SQLite queue.

4. 04

   Review

   Expose events through the API and operator dashboard.

04 / Validation

Evidence & scope

• Public C++ runtime, backend, dashboard, and architecture documentation.
• Replay-oriented testing and deterministic input workflows.
• An explicit event path spanning local processing, storage, and operator review.

05 / Perspective

Engineering takeaway

Reliable behavior around a model deserves its own tests. Replay makes it possible to reason about duplicate handling, queues, and event visibility without conflating those behaviors with model accuracy.