Skip to content
Back to projects
Artificial IntelligenceProduction

SensVX

AI-powered visual quality control for cable manufacturing

Role
Creator — AI & Product Engineering
Period
2024 — Present
PythonYOLOOpenCVComputer VisionNode.jsPostgreSQLEmbedded Hardware

Chapter 01

The problem

Cable manufacturing lines rely on manual visual inspection. Inspectors fatigue, defects slip through to customers, and every escaped defect means rework, downtime, and material loss. Generic off-the-shelf vision tools fail on the realities of a moving production line: variable lighting, motion blur, and defect types specific to cable extrusion.

Chapter 02

Research

The work started on the factory floor, not in a notebook: studying how defects actually appear on the line, which classes matter economically, and where cameras could physically mount. Dataset engineering followed — collecting images from real production runs, annotating defect classes manually, then cleaning and augmenting to survive real-world lighting variance.

Chapter 03

Architecture

  1. 01

    Physical Layer

    Production line, industrial cameras, and mounting/sensor infrastructure

  2. 02

    AI Layer

    YOLO-based detection and defect classification models with confidence scoring

  3. 03

    Backend

    Frame processing services and APIs feeding results downstream

  4. 04

    Database

    Persisted detections with full traceability per production run

  5. 05

    Dashboard & Analytics

    Live operator monitoring, trends, and alerts

Chapter 04

The solution

An end-to-end pipeline: camera capture → frame processing → image enhancement → YOLO detection → defect classification → confidence scoring → traceable storage → live dashboard. Low-confidence detections are flagged for human review instead of silently deciding — the system augments inspectors rather than pretending to replace them.

Chapter 05

Challenges

  • Maintaining detection reliability under changing line lighting and speed
  • Building a defect taxonomy that matches how the factory actually classifies waste
  • Keeping inference fast enough to be useful on a moving line
  • Integrating AI output into an operator workflow people trust

Results

  • Deployed against real production-line conditions in cable manufacturing
  • Continuous automated inspection replacing fatigue-limited manual checking
  • Full traceability: every detection stored, classified, and reviewable
  • Verified performance metrics to be published after production validation

Lessons Learned

  • Dataset quality beats model sophistication — most gains came from better data
  • Industrial AI is an integration problem as much as an ML problem
  • Operator trust is a design requirement, not an afterthought

Roadmap

  • Live AI inference demos
  • Multi-camera deployments
  • Real-time factory dashboard
  • Expanded defect taxonomy

Want the deeper technical story?

I'm happy to walk through architecture decisions, trade-offs, and demos.