Predictive ML & Scoring Models

Know what happens before it happens.

Predictive ML gives you a number instead of a hunch. Which lead will close. Which customer will churn. Which appointment will no-show. Which invoice will bounce. Same tech we built into RealZent's ShowUpAI — the model that kills the 40% no-show problem in Indian real estate.

Book Free Feasibility Call See ShowUpAI in Action

84% AUC in ShowUpAI

Auto-retraining built in

Explainable, not black-box

Site visit scoring · today ShowUpAI

3BHK, Bandra West

Visitor: Rajesh K · 3:00 PM

92%

Show up

2BHK, Powai

Visitor: Priya S · 5:30 PM

54%

Uncertain

4BHK, Andheri West

Visitor: Anand M · 7:00 PM

28%

Likely no-show

Retrained weekly · 47 features per visitor

What Is Predictive ML

A number that says what's about to happen.

Predictive ML is a model that learned patterns from your past data and produces a score for a new case. It doesn't write anything. It doesn't chat. It gives you a probability, a risk score, a forecast, or a ranking. You use that number to make a decision — send this customer a retention offer, book that appointment slot, restock this SKU, review that transaction.

People confuse this with generative AI because both are "AI". They're not the same tool. Predictive ML is older, more predictable, easier to explain, and often the right answer when the problem is "give me a number, not a paragraph".

Six problems predictive ML solves

No-show risk

Site visits, appointments

Churn probability

SaaS, subscription, telecom

Lead scoring

Sales pipeline priority

Demand forecasting

Inventory, staffing

Fraud detection

Payments, claims, refunds

Recommendations

Products, content, cross-sell

Proof of Work

ShowUpAI: how we killed the no-show problem

Real estate site visits have a 40% no-show rate in India. That's four wasted hours out of ten for every broker. We built the model that fixes it.

The Problem

Brokers wasted 40% of their day driving to empty flats.

The client says "yes, I'll come at 3". They don't. The broker has already blocked the slot, called the owner, told the security. Repeat 3–4 times a day. That's how the industry works.

What We Built

A model that scores every visit before it happens.

47 features per visitor — lead source, response time, WhatsApp read status, prior visit history, time of day, distance, property price band. XGBoost model outputs a probability of show-up. Below 40%, we auto-trigger OTP verification and broadcast reminder templates.

The Result

84% AUC. Show-up rate up 18–24 points per broker.

Not every visit is worth chasing. The model tells you which ones to double-confirm and which ones to just let go. Broker productivity per hour goes up because wasted trips go down.

The stack that shipped it

Model

XGBoost

Tabular data, fast inference

Feature store

Feast + Postgres

47 features per visitor

Serving

FastAPI + Docker

Sub-50ms scoring

Retraining

Weekly · MLflow

Automated drift alerts

See RealZent live

Six Patterns We've Shipped

The predictive problems we solve most often.

Each one has a live implementation in our own products.

No-Show & Cancellation Risk

Score every appointment, site visit, or booking. Real estate, clinics, salons, restaurants. Auto-trigger reminders below a threshold.

Live in: RealZent ShowUpAI

Customer Churn Prediction

Which subscriber, tenant, or client is likely to leave in the next 30/60/90 days. Trigger a retention play before they file the ticket.

Deployable to SaaS, subscription, telecom.

Lead Scoring & Prioritisation

Rank inbound leads by conversion probability. Route the top 20% to human sales, everyone else to a nurture flow. Cut cost-per-close.

Feeds directly into Zoho, HubSpot, Salesforce.

Demand & Sales Forecasting

Weekly, daily, or hourly demand by SKU, store, or channel. Reduces stockouts and overstock. Handles Indian holidays and monsoon patterns.

Prophet, NeuralProphet, or gradient-boosted trees.

Fraud & Anomaly Detection

Flag suspicious transactions, claims, refunds, or logins. Combines rule-based flags with an anomaly model. Fewer false positives than pure rules.

Sub-100ms scoring at the API layer.

Recommendation Engines

Products, articles, properties, courses. Collaborative filtering, content-based, or hybrid. RealZent uses one to match buyers to listings.

Handles cold start with content features.

Process & Pricing

Ship a scoring model in 6–10 weeks.

You get a baseline accuracy report before you commit to the full build.

Week 1

Data audit

We look at what you have. Volume, quality, features, target labels. Feasibility doc back to you.

Free

Weeks 2–3

Baseline model

Feature engineering + baseline XGBoost/LightGBM. You see the accuracy on holdout data.

₹1.5–3 L

Weeks 4–8

Production build

Feature store, API, monitoring, drift detection, retraining pipeline, dashboards, integration.

₹6–20 L

Ongoing

Operate

Monitoring, drift alerts, quarterly retraining, quality reviews. Model doesn't rot in production.

From ₹35k/mo

FAQ

Questions founders ask about ML.

Predictive ML is a machine learning model that guesses what happens next based on what happened before. Given your historical data — customer behaviour, transactions, appointments, sales — the model learns patterns and produces a score or number for a new case. Will this customer churn? Will this lead convert? Will this appointment no-show? What will sales be next month? The model tells you before it happens.

Generative AI writes text, images, or code. Predictive ML gives you a number or a probability. A generative AI drafts a newsletter. A predictive ML tells you which subscriber is 82% likely to unsubscribe next month. Different tools, different jobs. Predictive ML is older, cheaper to run, and usually gives you more explainable answers.

Prototype phase (data audit + baseline model + accuracy report) is ₹1.5–3 lakh. Full production build with feature store, retraining pipeline, monitoring, and integration into your app is ₹6–20 lakh. Ongoing operations start at ₹35,000/month and include monitoring, drift detection, and quarterly retraining. Feasibility call is free.

Depends on the problem. For churn or no-show scoring, we usually want 12+ months of historical data with at least 5,000–10,000 events. For demand forecasting, 2–3 years of daily or weekly data. For lead scoring, 1,000+ closed opportunities with outcomes. If you have less than this, we tell you before we take your money.

Depends on the signal in your data. RealZent's ShowUpAI predicts site-visit no-shows with 84% AUC. Churn models we've built land in the 78–89% AUC range. Lead scoring typically hits 0.75–0.85 AUC. We'll tell you the baseline accuracy after the prototype phase, before you commit to production build.

That's called data drift. Behaviour changes — economic cycle, seasonality, new product, new competitor — and the model starts predicting worse. We monitor drift automatically and trigger retraining. In the operations phase we retrain every 4–12 weeks depending on how fast your world moves. This is included in the monthly ops fee.

For tabular data: XGBoost, LightGBM, CatBoost, scikit-learn. For time series: Prophet, NeuralProphet, PyTorch Forecasting. For deep learning: PyTorch or TensorFlow depending on team preference. MLflow for experiment tracking, Feast for feature stores, Kubeflow or plain Airflow for pipelines. Deployment: AWS SageMaker, Azure ML, or bare Docker if simpler.

Yes. That's most of our clients. We handle the data science, MLOps, and monitoring. You just consume the score via an API or a dashboard. When we hand over, we give you an explainer doc so your team knows why the model says what it says — no black box.