Check out @TechCrunch's coverage of our $8 million series A round: .

RT @_WEEXIAO: It’s easy to fine-tune small models w/ RL to outperform foundation models on vertical tasks. We’re open sourcing Osmosis-App….

How did Temu become one of the fastest-growing e-commerce apps?. AI-powered personalization. Gamified UX. Multi-objective optimization. We break down the engine behind Temu’s engagement flywheel, and what other platforms can learn from it. 👇.

The H&M dataset is a staple in ML for fashion recs, massive scale, real user behavior, rich metadata. We show how to connect it to Shaped to build hybrid, sequential, personalized models in minutes. 📖

Vector DB ≠ personalization. Shaped handles what vector DBs don’t: deep user modeling, ranking, MLOps. Breakdown 👉

Can LLMs choose their own collaborative examples?. AdaptRec says yes - letting the model iteratively refine which user histories to learn from. 🧠 LLM-in-the-loop.📈 +18% HR@1 gains. Full breakdown 👉

Building “People to Follow” shouldn’t take a whole ML team. Shaped’s similar_users API lets you power smart user-to-user recs in one call 🤝. 📌 No complex infra.🧠 Learns from behavior + profiles.⚡ Real-time, relevant, ready to ship.Full post 👇.

Snowplow captures rich, event-level user data. Shaped turns it into real-time personalization. 🚀 Session-aware feeds.🔍 Behaviorally boosted search.🤖 No custom ML infra. Here’s how to connect them via Kinesis 👇.

The Shaped team had a blast at @netflix's RecSys Workshop 🎬.Top takeaways:.🧱 FMs are becoming infra.⚙️ Fewer models, more multitask.🧩 Domain grounding > raw LLMs.🔄 Meta-optimization is here.⚡ Efficiency still matters.Full recap 👇.

Two-Tower models are the backbone of real-time recommendations at scale. 🎯 Separate user/item towers.⚡ Fast ANN search for retrieval.🏗️ Key to multi-stage recsys pipelines.Why they work—and where they fall short 👇.📖

Criteo is still the benchmark for real-world ads + recsys:.📊 Sparse + dense features.🧠 Billions of rows.⚙️ Infra stress test.Why it still matters in 2025 👇.

Static recs miss what users want now. Sequential models fix that — modeling when users act, not just what they like. From N-Grams → Transformers → Generative Recs, here’s how modern systems capture real-time intent 👇.

Everyone wants a killer “For You” feed. But building one?.It’s ML, data engineering, infra, A/B testing hell. We break down what it really takes — and how platforms like Shaped make it 10x easier. 👇.

Still using Algolia for search and recommendations? You're paying the price in relevance and flexibility. Shaped is AI-native, unified, and built for real-time personalization. We did a full breakdown—here’s why it matters:

Meta's Jagged Flash Attention improves recommendation systems with up to 9x speedup and 22x memory reduction, outperforming dense flash attention by 3x and 53% in efficiency. Full write up here:

🚀 Introducing Shaped Value Modeling – full control over ranking optimization. Blend engagement, revenue, retention & more with a dynamic, interpretable framework. More here:

In our latest write up @bailuding & @Lunarmony introduce MoL, a retrieval model that learns similarity functions beyond dot products. It combines multiple embeddings for better accuracy in recommendations and QA while staying efficient.

MaskNet, introduced in 2021, boosts CTR prediction by using instance-guided multiplicative interactions. Its MaskBlock architecture still outperforms DeepFM and xDeepFM, improving AUC by up to 5.23%. Full article:

Still using notebooks & spreadsheets for analytics?. A fragmented stack slows ML teams down. Shaped Analytics unifies data, experiments & business impact—so you can iterate faster & optimize with confidence. Learn more:

Can data splitting impact recommender performance? . This study shows that different splitting strategies can drastically alter model rankings, challenging claims of state-of-the-art performance. Learn more:

Are traditional recommender systems outdated?. EmbSum leverages LLM-driven summarization to precompute rich user & content embeddings, outperforming UNBERT & MINER with fewer parameters. Handles 7,400+ tokens for deep personalization. 🤔👇.🔗