ECUBE ANALYTICS - Data Science And Machine Learning Business Consultancy And Education In Surat

Skin cancer Detection and classification using computer vision

(Computer Vision)
End-to-end pipeline for skin-lesion classification on the ISIC 2018 dermoscopy dataset. Compares a custom CNN against EfficientNet-B3 transfer learning, for both 7-class diagnosis and binary malignant/benign, with Grad-CAM explainability. PyTorch; deployed on Hugging Face. VIEW PROJECT

BedOps AI — Hospital Length-of-Stay Predictor

(MLOps (SageMaker + Cloud Run))
Predicts a hospital patient's Length-of-Stay category (4 bands) at admission, so ward managers can plan bed allocation and discharge ahead of time. Optuna-tuned LightGBM on 318k records, served from a live Amazon SageMaker Serverless Inference endpoint behind a Cloud Run UI. VIEW PROJECT

Echocardiography-Based LVEF Estimation with Explainable AI

(Computer Vision)

Estimates left ventricular ejection fraction (LVEF) from apical echocardiography videos using a 3D CNN trained on EchoNet-Dynamic (10,030 videos), with reduced-EF screening thresholds and 3D Grad-CAM explainability. Research prototype, not for clinical use. PyTorch; Streamlit app.

VIEW PROJECT

Predictive Maintenance: Failure Risk & Root-Cause Classification

(Machine Learning: Classification)

Two-stage ML system on 10,000 industrial sensor readings (AI4I 2020 dataset): an XGBoost model flags failure risk (PR-AUC 0.91), then a Random Forest model predicts the likely root cause among 5 failure modes (cross-validated macro-F1 0.95). SHAP-explainable. Streamlit app, deployed on GCP Cloud Run.

VIEW PROJECT

Credit Default Risk Prediction (Home Credit)

(Machine Learning: Classification)

Predicts loan-applicant default probability on the Kaggle Home Credit dataset (307K applicants, 7 linked tables). Aggregating the 5 auxiliary tables into about 150 features and stacking XGBoost, LightGBM and CatBoost lifted ROC-AUC from 0.66-0.73 (coursework baseline) to 0.788; the riskiest 10% of applicants contain 37.9% of defaulters. Gradio app on Google Cloud Run. Demo model, not a lending tool.

VIEW PROJECT

Stroke-Prediction

(Machine Larning: Classification)
Estimates stroke risk from routine clinical and lifestyle inputs (age, hypertension, heart disease, glucose, BMI, smoking status). Leakage-free, calibrated logistic regression: cross-validated AUC about 0.84 (age alone gives 0.83), with risk bands. Educational Streamlit app. VIEW PROJECT

DIABETES PREDICTION STREAMLIT APP

(Machine Learning: Classification)
Proof-of-concept tool that estimates the probability of diabetes from easily measured clinical features. Trained on a real but older, single-ethnicity dataset, so it is used for teaching and client demonstrations rather than diagnosis. Streamlit app. VIEW PROJECT

Predicting Probability Of Heart Attack

(Machine Learning: Classification)
Estimates 10-year heart-disease risk from routine inputs (age, blood pressure, cholesterol, glucose, BMI, smoking) on the Framingham data. Calibrated logistic regression, cross-validated AUC about 0.72, risk bands and a per-patient breakdown of what drives the estimate. On Hugging Face. VIEW PROJECT

SPAM_HAM-CLASSIFIER: E2E PROJECT WITH MODEL MONITORING AND MODEL DRIFT ANALYSIS

(Machine Learning: Classification)
Email spam/ham detector: FastAPI + Streamlit on Cloud Run. Body-only text cleaning and a TF-IDF logistic model trained on two public corpora (98-99% within a corpus, 66-88% across corpora), answering Uncertain or Too short instead of guessing. Airflow/Evidently retraining scaffold. VIEW PROJECT

Marketing Campaign Success Prediction

(End To End Machine Learning Projects)
Predicts whether a digital ad campaign will hit its conversion target, reframed as binary classification over budget, impressions, CTR, ad type and audience segment. XGBoost, ~91% test accuracy, custom UI. Deployed end-to-end on Google Cloud Run. VIEW PROJECT

Recommender System for Campaign CTR Prediction

(End To End Machine Learning Project)
End-to-end pipeline that predicts marketing-campaign click-through rate, reading training data from BigQuery and combining XGBoost regression with a hybrid recommendation approach. MLflow tracking, Docker, Cloud Build. Deployed on Google Cloud Run. VIEW PROJECT

Risk and Sentiment Intelligence Engine for Stock-Analyst Ratings Using PySpark and Azure Cognitive Services

(NLP+Azure Cognitive Services)
Text analytics for analyst-rating headlines: keyword-rule sentiment and risk labels with matched terms, an XGBoost approximation of the rules alongside, and optional Azure AI Language NER. PySpark feature prep, FastAPI, Streamlit UI; deployed on Google Cloud Run. VIEW PROJECT

PROPERTY PRICE PREDICTION FOR INDIAN METRO CITIES (REAL DATA)

(End To End Machine Learning Regression)

Predicts residential property prices across six Indian metros (Mumbai, Delhi, Kolkata, Chennai, Hyderabad, Bengaluru) from real 2022 data. A separate grid-search-tuned XGBoost model per city, with locality as a feature. Multi-page Flask app on Google Cloud Run.

VIEW PROJECT

A SUPERPOWERFUL GENERIC CLASSIFER (BINARY OR MULTICLASS)

(Machine Learning: Classification)
Upload any CSV or XLSX and build a classifier without writing code: automated EDA, 13 algorithms to choose from, slider-tuned hyperparameters, full evaluation reports, and seven diagnostic plots each with a short how-to-read note. 86 tests. Streamlit app. VIEW PROJECT

Multiagent Stock Market Technical And Fundamental Analysis

(Agentic AI)

A multi-agent system that answers natural-language stock questions using deterministic fundamental and technical analysis. Missing tickers are fetched on demand and cached as structured micro-indexes. Python 3.11, Alpha Vantage, RSI/SMA/EMA rules, Gradio UI.

VIEW PROJECT

LSTM-Based Stock Price Predictor

(Neural Networks)
Compares LSTM next-day AAPL forecasts with naive and moving-average baselines. Includes a returns-based LSTM and a retrain panel (window, units, dropout, epochs) with a validation split kept apart from the test period; says when a model is worse than 'tomorrow = today'. Streamlit on Hugging Face. VIEW PROJECT

VAR(Vector Autoregression)-Based Forecasting of Bitcoin, Gold, and S&P 500

(TIme_Series Forecasting)
Weekly Bitcoin, gold and S&P 500 forecast scenarios with simulated 50% and 90% bands from log-return models (random-walk bootstrap, VAR), weekly-close data refreshed through 2026, and a rolling backtest against a no-change forecast; CSV download. statsmodels, Gradio, on Hugging Face. VIEW PROJECT

Loan Eligibility Prediction Chatbot

(Machine Larning: Classification)
Explainable loan pre-screening demo on the public 614-row loan dataset. Logistic regression without gender or marital-status inputs, cross-validated AUC about 0.74 (credit history alone 0.68), exact per-input explanations and an Unlikely / Uncertain / Favourable band. Educational Streamlit app. VIEW PROJECT

Customer Churn (Banking) With SHAP Analysis

(Machine Learning: Classification)
Predicts which retail-banking customers are likely to churn, using a reproducible pipeline on real 10,000-customer data. XGBoost with SHAP explanations at both the segment and individual-customer level. Flask app with an interpretable, plot-driven UI. VIEW PROJECT

OCR + NER Legal Document Extraction App

(Computer Vision)
Extracts grantor/grantee, property address, APN, dates and more from scanned legal deeds. Runs fully offline (no cloud APIs): PaddleOCR plus a document-type-aware rule engine, scoring 67.6% strict field accuracy on six annotated development documents versus at most 3% for generic NER models. Gradio UI. VIEW PROJECT

Vision Transformer Based Gradio App For Empty Parking Lot Detection

(Computer Vision)

Classifies a whole parking-lot camera frame from the PKLot benchmark as mostly occupied or mostly empty (not a per-space detector). Fine-tuned ViT-B/16 reaches about 99% accuracy on days held out of training (split by calendar day to avoid near-duplicate time-lapse leakage), versus about 99% for a small CNN at the default threshold but with a weaker empty-lot recall. Generalisation to unseen lots is untested. Gradio app on Google Cloud Run.

VIEW PROJECT

PRECISION FARMING (SMART AGRICULTURE)

(Machine Learning: Classification)
Crop-recommendation tool that suggests the best crop to plant from seven soil and climate inputs: nitrogen, phosphorus, potassium, temperature, humidity, pH and rainfall. Multi-class Naive Bayes model. Streamlit app; second prize at a hackathon. VIEW PROJECT

Portfolio

Skin cancer Detection and classification using computer vision

BedOps AI — Hospital Length-of-Stay Predictor

Echocardiography-Based LVEF Estimation with Explainable AI

Predictive Maintenance: Failure Risk & Root-Cause Classification

Credit Default Risk Prediction (Home Credit)

Stroke-Prediction

DIABETES PREDICTION STREAMLIT APP

Predicting Probability Of Heart Attack

SPAM_HAM-CLASSIFIER: E2E PROJECT WITH MODEL MONITORING AND MODEL DRIFT ANALYSIS

Marketing Campaign Success Prediction

Recommender System for Campaign CTR Prediction

Risk and Sentiment Intelligence Engine for Stock-Analyst Ratings Using PySpark and Azure Cognitive Services

PROPERTY PRICE PREDICTION FOR INDIAN METRO CITIES (REAL DATA)

A SUPERPOWERFUL GENERIC CLASSIFER (BINARY OR MULTICLASS)

Multiagent Stock Market Technical And Fundamental Analysis

LSTM-Based Stock Price Predictor

VAR(Vector Autoregression)-Based Forecasting of Bitcoin, Gold, and S&P 500

Loan Eligibility Prediction Chatbot

Customer Churn (Banking) With SHAP Analysis

OCR + NER Legal Document Extraction App

Vision Transformer Based Gradio App For Empty Parking Lot Detection

PRECISION FARMING (SMART AGRICULTURE)