EPIC TECHNOLOGY · Computer Vision · Physical-World AI
Making real-world video AI dependable.
I led and built most of the software for a production multi-camera intelligence system under EPIC TECHNOLOGY, using AI-assisted development heavily. The difficult part was not choosing a detector: it was reliably connecting mixed-vendor cameras, knowing which source each frame belonged to, and keeping evidence and recovery consistent when media failed.
The real engineering problem
A model can be accurate on the wrong frame.
Camera streams can disconnect, freeze, repeat images, change routes or lose source identity. When a system treats “video URL reachable” as “fresh and correct camera,” the entire downstream AI chain becomes untrustworthy.
We needed practical computer vision on resource-constrained hardware and inexpensive mixed-vendor cameras, while still preserving operational event evidence. My direct work centered on the software and integration path; teammates handled camera and physical hardware installation.
System design
Four boundaries, one reliable event path.
An intentionally abstracted view, not a client deployment diagram.
Capture and qualify
Heterogeneous camera media must become identity-checked, fresh and consistently accessible before inference.
Perceive and interpret
Object detection, license-plate recognition, tracking and temporal rules turn pixels into bounded events.
Commit evidence
Asynchronous persistence, event histories and deduplication separate live decisions from slower record keeping.
Observe and recover
Feed health, source progress and bounded recovery keep a degraded camera from silently looking healthy.
Technical judgment
Decisions that mattered.
Software contracts mattered more than claiming a magic accuracy number.
Identity before inference
A feed with the wrong source or a frozen frame is a correctness failure even if the AI model confidently returns a detection.
Transport as a contract
A normalized media boundary lets detection and user-facing viewers consume heterogeneous vendor inputs without inheriting every proprietary behavior.
Temporal state over isolated frames
Events need direction, tracking, time windows and deduplication; raw frame detections are not operational answers.
Recovery without identity drift
Reconnect/restart paths should be bounded and source-qualified, with explicit states for interruption and recovery.
Implementation surfaces
Beyond model inference.
The work integrated YOLO-based perception, license-plate recognition, object tracking, temporal/event rules, standards-based video transport, role-aware dashboards, evidence capture, background persistence, monitoring and controlled recovery.
The resulting system has operated with multiple heterogeneous camera sources in a production setting. Exact client topology, vendor integration steps, accuracy, throughput, latency and uptime are not offered here as independently verified benchmarks.
What this demonstrates — and what it does not
End-to-end physical-world AI engineering.
Source validity, media acquisition, perception, temporal interpretation, persistence and recovery need a coherent owner. This project supports hands-on production video-system integration; it does not imply I installed the physical cameras, ran a factory MES/PLC integration or measured an unreported SLA.
Related work