Sanjib Samadder — The Storyteller

lead_prediction.py★ Featured · case study

Lead & Income Predictor: Gradient Boosting + Streamlit App

Predicts monthly leads, income and scrub cost for each partner and offer before money is committed, replacing gut-feel budgeting with a data-backed forecast. I compared seven models with cross-validation and deployed the winner as a live Streamlit app.

0.911CV R²

±0.017CV std

7Models compared

scikit-learnGradient BoostingXGBoostStackingStreamlit

Read case study → Code Live app

chest_xray_cnn.ipynbMachine Learning

COVID-19 Chest X-Ray Classification

Four neural network architectures (Dense, CNN, Deep CNN and LSTM) trained under identical conditions to separate COVID-19, viral pneumonia and normal X-rays.

→ A compact 3-block CNN won with 90.91% test accuracy and 17× fewer parameters than the Dense baseline. Adding depth made results worse.

TensorFlow / KerasCNNLSTM

Code

Cluster_Work.ipynbMachine Learning

Customer Segmentation: K-Means vs DBSCAN vs GMM

Behavioural segmentation of ~9,000 credit card customers, with engineered ratio features and skew correction, turned into risk and retention actions for a bank.

→ K-Means (k=5) beat both alternatives on silhouette, Davies-Bouldin and Calinski-Harabasz scores, with zero noise points.

scikit-learnPCAClustering

Code

basket_analysis.sqlData Analysis

Retail Sales & Basket Analysis

Cohort and market-basket analysis of 2.6M grocery transaction lines from 2,500 households, done entirely in PostgreSQL with CTEs and window functions.

→ After adjusting for basket size, pairs with lift above 1 fell from 96.5% to 44.9%. Bread + milk drops to near-chance (1.07). Indexing cut a lookup from 328 ms to 0.19 ms.

PostgreSQLWindow functionspandas

Code

lead_margin_forecast.xlsxData Analysis

Lead Margin Forecast & What-If Model

A 12-month income outlook built from 53 months of lead-buying history, with Base, Optimistic and Pessimistic scenarios, sensitivity grids, break-even analysis and a backtest.

→ Base case projects ~$122.7K income, with a ~$79K spread between the pessimistic and optimistic scenarios. Income is more sensitive to price than to cost.

ExcelForecastingWhat-If

Code Download Excel

credit_risk_analytics.xlsxData Analysis

Credit Risk Analytics & Scorecard

Default analysis of 32,581 real loans (Kaggle), built only with Excel formulas: an interactive dashboard, Weight of Evidence and IV, a points scorecard checked with KS and Gini, and an approval cut-off simulator.

→ Approving everyone loses $18.9M. A cut-off of 530 keeps 63% of applicants, cuts defaults from 21.8% to 6.2% and turns it into a $7.2M profit (Gini 73%).

ExcelCredit riskWoE / IVWhat-If

Code Download Excel

card_processing.pbixDashboard

Card Processing Performance Dashboard

A one-page Power BI report on a restaurant's merchant statements (Oct 2021 – Apr 2024): sales, transactions, fees and effective rate, with MTD, YTD and prior-year DAX measures on a star schema.

→ Fees are growing faster than sales (+2.17% vs +0.49% YTD). Keyed-in cards cost 3.69% against 2.28% when swiped.

Power BIDAXStar schema

Code View PDF

carbarn_master.datastudioDashboard

Carbarn Vehicle Import & Sales Dashboard

A live Google Data Studio control panel for a Japan → Australia used-vehicle importer: stock on the way, exporter instalments in JPY and AUD, shipping payables, customer balances and channel performance on one page.

→ 956 cars → A$24.0M turnover → A$6.92M profit (Jan 2022 – Mar 2026). Margin slid from 34.0% to 22.6% while average price rose, and sales moved from Facebook to Carsales and Marketplace.

Data StudioLooker StudioPayables tracking

Code Live report

lead_dashboard.xlsxDashboard

Lead Performance Dashboard

An interactive Excel dashboard for a lead-buying business. Filter by year, lead type and offer, and every KPI, ranking and chart updates across 53 months and 9,796 records.

→ 769K leads → $4.04M revenue → $2.16M income at a 53.5% margin. The top 5 of 77 partners generate 72% of income.

ExcelSUMIFSINDEX / MATCH

Code Download Excel

darwin_slp_fda.RmdStatistics

Darwin Sea-Level Pressure: Functional Data Analysis

75 years of climate data modelled as smooth curves with a Fourier basis, followed by functional PCA with VARIMAX rotation to find the dominant seasonal and ENSO modes.

→ Two components explain 79.71% of pressure variation. Manual λ selection improved RMS residual by 10.6%.

RfdaFPCA

Code

wine_clustering.RmdStatistics

Improving Clustering Quality with PCA

Used PCA loadings to identify and remove noise variables from the UCI Wine dataset before K-Means, then chose k using the elbow and silhouette methods.

→ Removing 3 mineral-derived noise variables raised cluster quality from 44.8% to 53.0%.

RPCAK-Means

Code

cancer_rates_fda.RmdStatistics

Australian Breast Cancer Rates: Functional Analysis (1921–2001)

B-spline smoothing with GCV-selected λ, functional descriptive statistics, rotated FPCA and functional regression testing whether rates in younger and older women move together.

→ PC1 explains 99.17% of variation (rise, peak, decline). β(t) ≈ 1.00 gives a significant positive association between age groups.

RB-splinesFunctional regression

Code

Sanjib Samadder
uses data to find the story behind it.

About me

Skills & stack

Experience

Research Assistant · SAFE Technologia Ltd, Chelmsford (UK)

Financial Data Analyst · Boardwalk Wealth (US real estate)

Freelance Data Analyst · Upwork, Top Rated Plus

Data & Business Analyst · Samsung Electronics, Walton BD, Omera LPG

Selected projects