Momentum AI: The Post-Purchase Precision Engine

Momentum AI: ML predicts the post-purchase 'Golden Window' for upsells, while local LLMs craft personalized messages. Maximize ROI with precision timing and private AI.

Comment

Momentum AI: The Post-Purchase Precision Engine

Inspiration

The idea for Momentum AI came from what we call the “Post-Purchase Paradox.”

A customer buys a laptop, and within minutes receives a generic “10% off” email for the exact same product. Most brands either:

  • re-engage customers far too early, creating spam,
  • or wait too long and lose customer intent entirely.

We wanted to build a system that treats timing as a variable, not a constant.

The core inspiration was the concept of Optimal Re-engagement Windows — identifying the Golden Window where:

  • customer intent,
  • product utility,
  • and purchase readiness

all intersect.

What it does

Momentum AI is a post-purchase timing and personalization engine designed to convert one-time buyers into repeat customers.

The platform operates in three stages:

1. Predictive Timing

A Machine Learning model powered by Random Forest Regression analyzes:

  • product category,
  • customer segment,
  • order value,
  • and behavioral patterns

to determine the optimal delay before re-engagement.

The predicted wait period is:

$$ \hat{y} = \sum_{i=1}^{n} Tree_i(X) $$

Where:

  • (X) represents the encoded feature vector,
  • and each decision tree contributes to the final prediction.

The goal is to determine the ideal:

$$ d_{wait} $$

before sending the next upsell or retention message.

2. Generative Personalization

Once the timing is determined, a local Large Language Model running through Ollama generates:

  • hyper-personalized follow-up messages,
  • contextual upsell recommendations,
  • and customer-specific engagement copy.

This ensures every interaction feels intentional rather than automated.

3. Customer Intelligence Dashboard

Momentum AI also provides a premium analytics dashboard featuring:

  • Customer Personas,
  • Purchase Timelines,
  • Re-engagement Predictions,
  • and behavioral insights.

This allows brands to understand not only what the AI predicts, but why it predicts it.

How we built it

We built Momentum AI using a high-performance local-first Private AI architecture.

The Brain — Predictive AI Layer

Built using:

  • Python
  • FastAPI
  • scikit-learn

The system uses a RandomForestRegressor trained on synthetic e-commerce behavior data.

Feature vectors include:

  • encoded customer segments,
  • product categories,
  • purchase values,
  • and transaction metadata.

The Voice — Generative AI Layer

For inference, we used:

  • Ollama
  • running the lfm2.5-thinking model locally.

This architecture guarantees:

  • zero external API dependency,
  • complete data privacy,
  • and enterprise-safe deployment.

The Face — Frontend Experience

The frontend was intentionally built as a:

  • Zero-Framework Single Page Application

using:

  • Vanilla HTML,
  • CSS,
  • and JavaScript.

We relied heavily on:

  • CSS Grid,
  • Flexbox,
  • transitions,
  • and Chart.js

to create a premium SaaS-style interface without React or Tailwind overhead.

The Data Layer

We used:

  • SQLite
  • with 500+ synthetic customer purchase records

to simulate realistic e-commerce usage patterns and model training conditions.

Challenges we ran into

1. The Latency Gap

Running a local “thinking model” introduces noticeable inference latency.

We solved this by:

  • implementing asynchronous background tasks in FastAPI,
  • and designing UI “thinking states” using animated typing indicators.

This preserved perceived responsiveness while maintaining local inference quality.

2. Categorical Complexity

Human customer types such as:

  • VIP,
  • Returning,
  • New Customer

cannot be treated as ordinal numeric values.

To solve this, we implemented One-Hot Encoding:

$$ x_{segment} \in { [1,0,0], [0,1,0], [0,0,1] } $$

This prevented the model from incorrectly assuming hierarchical relationships between customer classes.

3. CSS-Only UI Engineering

Creating:

  • slide-in customer intelligence panels,
  • animated phone mockups,
  • and reactive dashboard components

using only Vanilla CSS required extremely precise:

  • positioning,
  • transition timing,
  • and state management logic.

Achieving a premium modern UI without frontend frameworks was significantly harder than expected.

Accomplishments that we're proud of

Zero External Dependencies

The entire application runs:

  • without API keys,
  • without cloud inference,
  • and without third-party AI services.

This makes Momentum AI:

  • fully private,
  • enterprise-ready,
  • and deployable in restricted environments.

Dual-AI Architecture

We successfully combined:

  • Predictive AI (scikit-learn)
  • and Generative AI (Ollama)

inside a single seamless workflow.

This created a unified intelligence pipeline rather than isolated AI features.

Premium UX Without Frameworks

We achieved:

  • a modern SaaS aesthetic,
  • reactive interactions,
  • and smooth transitions

using only:

  • Vanilla JavaScript,
  • native CSS,
  • and browser APIs.

No React. No Tailwind. No build tools.

What we learned

The biggest lesson was:

Timing matters more than discount size.

A:

  • 5% discount at the correct moment

outperforms:

  • a 50% discount delivered at the wrong time.

We also learned that:

  • Local AI has matured enough for interactive SaaS applications,
  • provided the user experience is engineered around model thinking time and asynchronous execution.

What's next for Momentum AI: The Post-Purchase Precision Engine

Our next goal is transforming Momentum AI into an:

Autonomous Growth Agent

The roadmap includes:

Self-Optimizing Reinforcement Loops

We plan to integrate:

  • Reinforcement Learning
  • with live conversion feedback

so the system continuously improves its timing predictions based on:

  • successful upsells,
  • customer engagement,
  • and conversion outcomes.

Multi-Channel Execution

Future versions will support automated engagement across:

  • WhatsApp,
  • Email,
  • SMS,
  • and push notifications.

The system will automatically trigger communication when the:

  • “Golden Window” opens.

Enterprise Expansion

Additional future capabilities include:

  • real-time customer lifetime value prediction,
  • churn forecasting,
  • campaign ROI attribution,
  • and AI-driven retention automation.

The long-term vision is a fully autonomous post-purchase intelligence platform that continuously optimizes customer retention without manual intervention.

Built With

Share this project:

Updates