Portfolio DemoSimulated DataAI-assisted

AMC Filter Analytics — Methodology Demo

Three models for AMC (Airborne Molecular Contamination) filter analytics, built on public activated-carbon and air-quality datasets. Methodology demo only — all data and parameters are simulated or sourced from public literature.

Try a prediction See MLOps

Breakthrough Time Prediction

Thomas R² 0.995 · XGBoost R² 0.84

Predict when a fixed-bed adsorber reaches 10% breakthrough (C/C0 = 0.1) given flow, concentration, and bed parameters. Thomas analytical baseline + XGBoost observable-feature regressor.

Regeneration Cycle Degradation

LSTM MAE 0.46 · CatBoost MAE 0.68 mg/g

Forecast remaining adsorption capacity after N regeneration cycles from a rolling history window. CatBoost deployed in API; LSTM benchmarked in training repo.

Batch Quality Anomaly Detection

Precision / Recall / F1 ≈ 0.84

Flag incoming activated-carbon batches likely to yield poor regeneration from four QC features (BET, porosity, ash, moisture). Isolation Forest with intentionally realistic overlap.

About this demo

A portfolio project demonstrating three AMC filter analytics models — breakthrough-curve prediction, regeneration-cycle degradation, and batch quality scoring — plus the MLOps shape around them (training → metrics → drift → scheduled retrain). The point is to show the methodology and engineering end-to-end, inspectable from training code through to the deployed UI.

Every response from the API carries data_source: "simulated". All data and parameters are either synthetically generated or sourced from public literature and public APIs.

Stack

Frontend: Next.js 16 + shadcn/ui + Recharts
API: FastAPI via shared dashai-api gateway (/amc)
Training: XGBoost · CatBoost · PyTorch · scikit-learn
Deploy: Vercel (frontend) · Render (backend, Singapore)