Recommendations that actually understand users.

Cold-start & sparsity

Pure ID-based embeddings need historical interactions. New games, new posts, and new users don’t get meaningful recommendations until it’s too late.

Static user profiles

Most user towers are built on slowly moving aggregates. They don’t know if the user is currently in a “quick-snack scroll” vs “deep research” mode.

Opaque ranking logic

Even when metrics go up, it’s hard to explain why a recommendation appeared, or to prove it satisfies safety and compliance constraints.

1. LLM-generated tower embeddings

We enrich both user and item towers with LLM-generated embeddings that encode deep semantic meaning from text and metadata—not just IDs.

• Better recall@K and semantic matching
• Cross-domain generalization (e.g., game ↔ post ↔ video)
• Stronger cold-start for low-history entities

2. Context-aware user representations

LLMs summarize recent sessions, like “competitive arcade games + short comics + FIFA soccer content”, and turn them into contextual embeddings.

• Session-based ranking for feeds and games
• Alignment with “Good Visit” or intent labels
• Dynamic personalization over static personas

3. Hybrid retrieval + LLM reranking

Two-tower retrieval handles millions of items in tens of milliseconds; a batched LLM reranker refines the top candidates with rich reasoning.

• Two-tower recall in 10–20 ms
• LLM rerank in 50–150 ms (batched)
• 10–30% gains in nDCG/precision in our target scenarios

4. LLM-generated feature layer

Topic clusters, quality scores, toxicity signals, difficulty tiers, and item summaries are all produced via LLMs and fed into both towers.

• Richer features without manual feature engineering
• Shared embeddings across text, image, and game assets
• Cheaper at scale than custom NLP stacks

5. Cold-start by design

For new content, we generate item embeddings directly from text and multimodal metadata, enabling meaningful similarity and ranking on day one.

• Zero-shot similarity (“this game is like Arcade”)
• New post/game/product performance from hour 0
• Long-tail content no longer starved

6. LLM-generated synthetic data

Two-tower models are data-hungry. We use LLMs to generate realistic positives/negatives and rare behavior patterns to stabilize training.

• Expanded coverage for edge segments
• Reduced sampling bias
• Improved generalization in sparse regimes

7. LLM-guided hard negative mining

We ask the LLM to identify “semantically close but wrong” items and feed them as hard negatives into contrastive two-tower training.

• Higher discriminative power in the embedding space
• Improved recall@K where it matters
• More stable multi-task training

8. Knowledge, safety & explanation layer

On top of tower embeddings, an LLM reasons about why an item matches, enforces safety rules, and generates auditor-friendly explanations.

• Policy-aware recommendation gating
• Human-readable rationales for trust, safety & legal teams
• Faster debugging for PMs and engineers

Social & content feeds

• communities, forums, & discussion feeds
• Session-aware ranking for returning users
• Explainable controls for policy & mod teams

Games & interactive experiences

• Personalized game discovery and difficulty tiers
• Context-aware matchmaking & quest suggestions
• Cold-start for new titles and in-game items

Marketplaces & ads

• Semantic matching between users, items, and intents
• LLM-driven enrichment of sparse product metadata
• Safety-aware ranking and explanations for advertisers

What’s covered

• LLM-generated embeddings feeding user & item towers
• Dynamic user representations via LLM session summaries
• Hybrid two-tower retrieval + LLM reranking
• LLM-generated synthetic data & hard negative mining
• LLM reasoning layer for safety, compliance, and explanation

Why it matters

As recommendation stacks converge on similar architectures, IP around how LLMs integrate with towers, synthetic data, and explainability will matter—for partnerships, defensibility, and long-term product strategy.

Note: Non-provisional is prepared with counsel.