Signal

Explainable lead scoring: train a logistic-regression model in the browser, evaluate it properly, and hand reps a ranked call list with reasons.

Live demo →

The problem

Sales teams have more leads than time. Calling in order of arrival wastes most calls, and black-box scores that reps don't trust get ignored. A good lead score needs to be accurate, honestly evaluated, and explainable.

What it does

• Trains a logistic regression in the browser on 1,500 sample leads, or on any CSV with a 0/1 converted column.
• Evaluates it only on a held-out test set: AUC, ROC curve, and a cumulative gains chart.
• Translates the model into business terms: "call the top 20% of leads and you find 55% of the buyers, 2.8× better than random."
• Shows a confusion matrix for the chosen call-list size: found, missed, wasted calls, saved calls.
• Charts standardized feature weights (what drives conversion), with odds ratios on hover.
• Gives every lead the top reasons behind its score.
• Lets you toggle features on and off and adjust regularization to see the effect.

How it works

• Features: text columns are one-hot encoded and numeric columns are standardized using training-set statistics only, so nothing leaks from the test set. ID-like columns (lead_id, email, …) are skipped automatically.
• Training: batch gradient descent on log-loss with L2 regularization, 400 epochs, on a seeded random train/test split.
• AUC: computed with the Mann–Whitney rank formulation, with ties handled.
• Explanations: for a linear model, each feature's contribution to the log-odds is exactly weight × value, so the reasons shown are the actual drivers of that score.

The sample leads are generated from a known logistic process plus noise, so the model's job is to recover that signal.

Run it locally

No build step and no dependencies. Serve the folder with any static server:

python3 -m http.server 8000
# open http://localhost:8000

Project structure

index.html   page layout and copy
app.js       data generation, model, and rendering
app.css      page-specific styles
kit.js       shared helpers: seeded RNG, CSV import/export, SVG charts, tooltips
kit.css      shared UI kit (light + dark mode)

Charts are hand-built SVG with no charting library. Everything runs client-side, so uploaded files never leave the browser.

Data

The bundled sample data is synthetic, generated from a seeded random process so it is reproducible. It does not describe any real company or customer.

---

Built by Darshan Patil · LinkedIn