Sergey Kopanev - Entrepreneur & Systems Architect
Go Back

User Intent Prediction: The Plan

I had a perfect plan. And I truly believed it. That’s the embarrassing part.

Because nothing else worked. Not analytics, not surveys, not “best practices.” If I wanted to understand intent, ML felt like the only weapon left.

I was going to solve acquisition in high-friction health funnels using machine learning. I was going to stop guessing and start predicting. I was going to turn chaotic user behavior into pure revenue.

Here is exactly what I thought would happen.

The Problem

Health & fitness funnels are chaos.

  • Length: 30–60 screens.
  • Conversion: 1.5–2.5%.
  • Behavior: Irrational.

People jump from Screen 3 to Screen 13, then back to Screen 2 like they’re speedrunning a game. Traditional analytics misses intent completely. You see what they did, but never why.

The Business Goal

I didn’t want to build a cool model. I wanted to stop wasting traffic.

  1. Identify who will buy and who won’t.
  2. Adapt the funnel in real time.
  3. Raise acquisition.

If I could predict intent at Screen 10, I could change the offer at Screen 60.

The Hypotheses

I believed machine learning could solve this if I:

  1. Encoded behavior into embeddings (User2Vec).
  2. Compressed the data to find the “signal” (PCA/UMAP).
  3. Predicted purchases from patterns (Deep Learning).
  4. Modeled timing (dwell time, idle gaps).
  5. Personalized discounts via bandits.
  6. Deployed a real-time model (< 500ms).

I genuinely believed this stack would outperform anything hand-crafted.

Why I Thought This Would Work

1. Behavior looks like language. A user session is just a sentence of screens. If Transformers can understand English, they should understand funnels.

2. 44,000 users is “enough data.” In my mind, 44K sequences was a goldmine.

3. Deep Learning finds patterns humans miss. I thought more complexity would automatically mean more insight.

The Expected Metrics

I set my targets. They seemed reasonable.

  • Prediction Quality: 0.80–0.90 (Strong prediction)
  • Inference: < 500ms (Real-time)
  • Uplift: +10–20% conversion
  • Result: One unified intent signal.

Looking back, these numbers were pure fantasy.

The Build Plan

  1. Create embeddings.
  2. Test separation.
  3. Compress (PCA, t-SNE, UMAP).
  4. Build purchase predictor.
  5. Add dwell + idle + time-series features.
  6. Build motivation model.
  7. Add bandits.
  8. Convert to ONNX.
  9. Benchmark latency.
  10. Produce final intent engine.

The Naive Belief

I believed that end-to-end learning would “understand users.” I believed that embeddings would encode intent. I believed that deep learning would beat simple features. I believed that funnels behave logically.

I wasn’t just wrong — I was delusional. Funnels don’t behave. Users don’t behave.

The Cliffhanger

The plan looked brilliant on paper. Reality set it on fire.

Next: User Intent Prediction #1: 32 Dimensions to 8. Everything Broke.