Sage Nord
Sage Nord Lab
Deep research and investigations into software engineering and AI.
- Why Uber built one gateway for 60+ internal LLM use cases instead of letting every team integrate separately Uber had over 60 internal LLM use cases with each team building its own integration. The GenAI Gateway centralized vendor access, security, and cost tracking behind an OpenAI-compatible API, and it also shows what that centralization costs.
- How Discord fixed hot partitions by moving trillions of messages off Cassandra Discord's Cassandra cluster grew to 177 nodes and kept paging on-call for unpredictable latency caused by 'hot partitions.' The fix wasn't just a database swap to ScyllaDB, it was a coalescing layer in front of it.
- How Shopify sharded Shop app's Rails backend without stopping the world Shopify's engineering team migrated the Shop app's MySQL backend to Vitess to fix multi-week schema migrations and capacity throttling. The real lesson isn't the sharding key, it's the query verifiers that made the migration safe.
- Why on-device LLM assistants have felt slow, and how NPUs fix it A Peking University ASPLOS 2025 paper shows NPU offloading can cut mobile LLM prefill latency dramatically. Here's why that matters for anyone building on-device AI features on iOS or Android.
- EU AI Act Article 50 is now in force: what chatbots, generative AI, and deepfake features must do Article 50 of the EU AI Act took effect on 2 August 2026, requiring AI transparency disclosures and machine-readable watermarking. Here's what it actually requires, verified against the official text, and what it means for teams shipping AI features.
- What 86,726 broken code samples tell us about trusting AI-generated code A new IEEE ISSRE 2026 study quantifies how LLM-generated code actually fails across languages, and shows that self-correction is far less reliable than assumed. Here's what it means for teams shipping AI-assisted code.