Hoshang Sheth logoHoshang Sheth
All Services
Service Deep Dive

Machine Learning Systems.

Predict what's coming. Act before the damage hits.

Book a Discovery Call Ask a question
At a glance
  • Timeline · Custom estimate
  • Investment · ₹60,000 - ₹95,000
  • Model · Fixed scope, fixed price
  • Ownership · Full source, no lock-in
Weeks early
you see it coming weeks before it costs you
Data-driven
decisions backed by your own data, not gut feel
85%+ AUC
on Recovia, a founder-built fintech product

Figures from founder-built products and published benchmarks. Client results vary by scope, data quality, and use case.

The Problem

Why this matters right now.

Most business decisions are made after the damage is done. You find out a customer churned after they cancelled. You discover a stockout after losing the sale. You detect fraud during reconciliation, not before it clears. ML systems flip that timing.

The Approach

How I build this.

  1. 1Before any modelling starts, we agree on what decision the prediction is supposed to drive otherwise the model sits unused.
  2. 2We work with your actual operational data, not synthetic examples as the model learns from what your business has really seen.
  3. 3Every prediction comes with an explanation of why the model scored it that way, no black boxes, no blind trust.
  4. 4Predictions surface in a dashboard or alert your team already checks - not a separate tool nobody opens.
  5. 5The model gets retrained as your business changes so it doesn't quietly degrade over time.
What you get

Deliverables

  • Churn prediction model with 30-60 day advance warning
  • Demand and inventory forecasting from your operational data
  • Fraud and anomaly detection with real-time alerts
  • Lead scoring model ranked by likelihood to convert
Under the hood

Technology stack

PyTorch / TensorFlowScikit-learnSHAPPostgreSQLFastAPIReact

PyTorch for deep learning tasks - sequence models, embeddings, neural networks. Scikit-learn for classical problems where simpler models generalise better. SHAP keeps every prediction explainable regardless of which approach is used.

Case Study

Recovia - explainable loan default risk

Portfolio demonstration
Problem

Lenders relying on heuristics to identify high-risk borrowers had delayed recovery actions and poor prioritisation. Manual scoring didn’t scale.

Solution

XGBoost classifier for borrower default risk + SHAP-based explainability (so nothing is a black box) + KMeans customer segmentation + automated PDF risk reports - all served via a FastAPI + React dashboard.

XGBoostSHAPKMeansFastAPIReact
Live GitHub
Investment

₹60,000 - ₹95,000

Fixed scope, fixed price

Every ML engagement starts with a data-readiness assessment on the discovery call. Where you land in the range depends on data quality, model complexity, and how predictions get delivered into your workflow.

Every project is quoted as a single fixed price once scoped. The range above reflects how much project size varies - not variable billing within a project.

Typical Delivery
Custom estimate
Payment Terms
50% to begin · 20% at prototype · 30% before handover
Frequently Asked

Questions specific to machine learning systems.

It depends on the problem, but as a rule of thumb: 5,000+ labelled records for classification, 12+ months of history for forecasting. If data is thin, we can start with a heuristic model and upgrade later.
Yes. Every model ships with a SHAP-based explanation panel showing which factors pushed a score up or down for any individual prediction.
Every ML system I ship includes a monitoring layer: input distribution drift alerts, prediction distribution shift, and a scheduled retraining pipeline.
Yes. The model is only useful if predictions become actions - automated emails, CRM alerts, priority reordering. That plumbing is part of the engagement.
Most real business data isn't and that's normal. Data cleaning and preparation is part of the engagement, not a prerequisite for starting. If the data turns out to be genuinely too thin or too messy to build a reliable model, I'll tell you that on the discovery call rather than take the project and deliver something that doesn't work.
Next Step

Ready to scope machine learning systems for your business?

Bring your business problem, not a dataset. We'll work out on the call whether your data is ready and what a model could realistically predict.

Book a Discovery Call See all services