Skip to content
All work
04Predictive ML · Web App2025

MachineInsight Pro — Predictive Maintenance

Telemetry in, failure-risk out — an IBM Watson AutoAI LightGBM model served through a Vercel serverless proxy.

Impact

99.4%

F1-weighted accuracy

Milliseconds

Telemetry to prediction

Live

Deployed on Vercel

Technology

Core
HTML5Tailwind CSSChart.jsGSAP
Infrastructure
Vercel Serverless FunctionsVercel
AI / ML
IBM Watson AutoAILightGBMAutomated HPO + feature engineering

Key decisions

  1. 01

    A serverless proxy to hide the model key

    Calling Watson directly from the browser would leak the API key and break on CORS. A Vercel function injects the credential server-side, so the secret never reaches the client.

  2. 02

    AutoAI to find the model

    Watson AutoAI ranked multiple pipelines and selected a LightGBM classifier with engineered features and tuned hyperparameters, instead of me hand-picking a model.

  3. 03

    A telemetry-driven, real-time UI

    The UI takes the five telemetry signals and shows the failure-risk prediction immediately with gauges and charts, so an operator reads it at a glance.

Architecture

Architectureflow
  1. 01

    Telemetry input

    Torque, speed, temperature, tool wear, quality type

  2. 02

    Vercel serverless /api/predict

    Injects IBM credentials server-side, forwards the request

  3. 03

    IBM Watson AutoAI

    LightGBM classifier, automated feature engineering + HPO

    • LightGBM
  4. 04

    Prediction

    Failure-risk returned in milliseconds

  5. 05

    Dashboard

    Glass-morphism UI with gauges and animations

    • Chart.js
    • GSAP

The problem

Industrial maintenance is usually reactive — you fix a machine after it fails, and the downtime is expensive. Predicting failure from live telemetry needs a trained model, but calling it straight from the browser would expose the model's credentials and hit CORS limits.

How it works

I used IBM Watson AutoAI to generate, rank, and tune models on machine telemetry; it selected a LightGBM classifier with automated feature engineering and hyperparameter optimisation, reaching 99.4% F1-weighted accuracy. The browser sends five telemetry signals (torque, speed, temperature, tool wear, quality type) to a Vercel serverless function at /api/predict, which injects the IBM credentials server-side and forwards the request to the Watson AutoAI endpoint. The result comes back to a glass-morphism UI with Chart.js gauges and GSAP animations.