← Back to Work

AI-Powered Recommender Systems

Building ML-Driven Content Discovery at Scale

TIMELINE 2019–Present
ROLE Senior PM
TEAM 8–12 People

Overview

In 2019, I joined Seznam.cz as Senior Product Manager responsible for recommender systems. Seznam is the largest Czech search engine and digital platform with millions of daily active users. My mission: own the product strategy for recommendation engines powering content discovery across the platform.

This has been my most complex product challenge. The recommender system serves 50M+ requests daily, powers real-time personalization for logged-in users, and directly impacts engagement metrics like time on site and user retention. The scale is massive, complexity is high, but connecting people with relevant content is what makes it compelling.

The Challenge: Balancing massive scale (50M+ daily requests) with real-time semantic relevance, while avoiding filter bubbles and ensuring content diversity.

The Problem: Discovery at Scale

"What We Solved"

Seznam's challenge was fundamentally about relevance and scale. Every day, thousands of articles are published. How do you surface the most relevant content to each user in milliseconds?

  • Information Overload: Too much content, limited attention span.
  • Cold Start: Recommending to new users without history.
  • Diversity vs. Relevance: Avoiding filter bubbles while staying personal.
  • Latency: <100ms response time required for UX.

The opportunity was to build a system that ingests user behavior, understands content semantics via NLP, and generates highly relevant recommendations at massive scale.

Approach & Strategy

Three Core Pillars

1. Hybrid Recommendation Architecture

What it was: Collaboration of collaborative filtering (user-to-user), content-based filtering (semantics), and knowledge graphs.

Why it mattered: No single algorithm works for everyone. Collaborative is great for engaged users; content-based solves cold start; knowledge graphs ensure diversity.

2. Real-Time Personalization

What it was: Profiles updated instantly on clicks/reads. Recommendations recomputed in real-time. Extensive A/B testing.

Why it mattered: Freshness is critical. If a user reads a topic now, their intent has changed. Real-time signals beat overnight batch processes.

3. Responsible AI & Diversity

What it was: Explicit optimization for diversity and serendipity. Guardrails against filter bubbles and over-concentration.

Why it mattered: Personalization unchecked is dangerous. We balanced "what users want" with "diverse perspectives" to support healthy info consumption.

Technical Architecture

CORE STACK
Python, Jupyter, Spark, Hadoop
ML & MODELS
TensorFlow, Vowpal, LLMs (Small-E-Czech, ChatGPT)
INFRA & OPS
Gitlab, Docker, AirFlow, SQL
ANALYTICS
Tableau, Custom Dashboards

Execution: Timeline

Phase 1 (2019 – 2020): Baseline

Audited existing systems. Built data pipelines. Launched first ML models on homepage (beta). 8-12% lift in CTR.

Phase 2 (2020 – 2021): Scaling

Real-time personalization. Expanded to all content types. 100% rollout. 15-20% engagement improvement.

Phase 3 (2021 – Present): Advanced

Deep Learning models (NCF). Diversity constraints. Reranking for freshness. 20-25% overall engagement boost.

Key Metrics & Impact

50M+

Daily Requests

<150ms

p99 Latency

+25%

Engagement

50+

Feature NPS

*Focus shifted from "Shallow Clicks" to "Deep Engagement" (Time spent reading/watching).

Lessons Learned

1. Perfect Models Don't Exist

Offline metrics (AUC) don't always match online reality. We optimized for fast iteration and online A/B testing over perfect offline accuracy.

2. Data Quality is Foundation

Garbage data = garbage recommendations. We now budget 40-50% of effort on data infrastructure rather than model complexity.

3. Explainability Matters

Black boxes destroy trust. Users need to understand "Why this?". Adding explainability features increased user trust and satisfaction.

4. Continuous Monitoring

ML systems are fragile. Data shifts happen. We built robust monitoring to detect when models degrade before users notice.

Building ML systems at scale?

Balancing model performance with responsible AI is the challenge of our time. Let's discuss how to turn data into value responsibly.

Let's Discuss