Jiaxing Hu

I wrote an article on how we translate user-generated content at scale at Mercari. It summarizes two years of work with various models, LLMs but not only, and how cost evolved, reducing it 100x over the period. I also go over non-AI features and user experience. https://lnkd.in/gbpQAQmX

39

2 Comments

Huge thanks to Zhe Jia, Shuai Wang, Aditi Bamba, and Martin Wang from DoorDash for publishing this incredibly practical breakdown of their RAG implementation for Dasher support. This post saved me countless hours of trial and error. The two-tier guardrail approach is what really clicked for me: * Run a cheap semantic similarity check first (sliding window comparison between response and KB articles) * Only if that fails, escalate to the full LLM evaluator Result: 90% fewer hallucinations, 99% fewer compliance issues, without killing your latency budget When you're actually shipping RAG to users, this kind of practical filtering makes all the difference. Everyone talks about prompt engineering, but nobody talks enough about how to catch bad outputs before users see them. Their broader setup—quality monitoring across 5 dimensions, regression test suites, the whole LLM Judge framework—gives you a real blueprint beyond the basic tutorials. You can't fix what you're not measuring. The article is https://lnkd.in/dbC9BG7a

2

Localization is more than translation. Pendo.io just rolled out our first localized experience—in Japanese 🇯🇵—designed for Japan’s fast-growing, AI-driven tech market. Go Pendo.io Japan ! From login screens to product areas, we’re bringing product-led growth to users in their language. Learn how we're doing it—and why Japan was first! 👇

2

Developing a large ecosystem with both open- and closed-source components comes with some interesting challenges. Back when we were bootstrapping CKEditor, my purist mindset pushed us to develop each of our 30+ packages in separate repositories. The goal was to sort of "force" higher modularity and more responsible architecture for every component. As you can probably guess 😅, it did not scale. Later, we consolidated most of the packages into two monorepos – one open source, one closed source – while keeping a few granular packages separate where it still made sense. This helped a lot, but it didn’t fully solve the problem. We were still working across multiple repositories and kept running into issues like dependency coordination and lockfiles. Recently, we took another step forward. We migrated to pnpm and redesigned our overall approach. In this post, Filip Sobol explains the new setup implemented by the CKEditor 5 Platform team and what we learned along the way: https://lnkd.in/eHS28qxe

23

4 Comments

Over the past year, my team and I built Pinterest's Model Context Protocol (MCP) ecosystem from the ground up: a central registry, a growing fleet of domain-specific MCP servers, and production integrations across our IDEs, internal chat surfaces, and AI agents. MCP is an open standard that gives large language models a unified way to talk to tools and data sources. At Pinterest, we used it as the substrate for AI agents that can safely automate real engineering tasks, from querying Presto data on demand to diagnosing Spark job failures and surfacing institutional knowledge. A few things I'm particularly proud of: * We designed for security from day one: two-layer auth (end-user JWTs + mesh identities), business-group-based access gating, and mandatory human-in-the-loop for sensitive actions. * We made it easy for any team to ship a new MCP server by creating a unified deployment pipeline that handles all the infrastructure, so domain experts just write business logic. * The ecosystem has scaled to 66,000+ invocations per month across 844 monthly active users, saving an estimated 7,000 engineering hours per month. This was a cross-functional effort. Thank you to everyone on the Agent Foundations, Security Engineering, and Traffic Engineering teams who made this possible, and to our engineering sponsors for their support and guidance. Full write-up on the Pinterest Engineering Blog: https://lnkd.in/e-mH26Bw

146

7 Comments

🚀 New Update: Dynamic Image Transformation for Amazon CloudFront We’re excited to roll out the latest version of Dynamic Image Transformation for Amazon CloudFront, making it easier than ever for teams to deliver fast, optimized, device-aware images through Amazon CloudFront's global CDN. This release introduces: ✨ Automatic image optimizations, tailoring image format, compression, and size to each user's device ✨ Support for delivering large, high-resolution images without compromising performance ✨ Support for external image origins, including popular CMS platforms ✨ An intuitive configuration experience to get up and running faster If you’re at AWS re:Invent this week, stop by the Edge & Networking area on the Expo floor to see a live demo and walkthrough of DIT. You can also explore other new CloudFront and Edge capabilities launching this week! https://lnkd.in/guc3pqhm

18

An agent has no consistent definition. You can think of an agent as an autonomous or semi-autonomous system that can take actions on behalf of the user in a given environment, state, and make decisions or take actions that can accomplish certain tasks along a given time-frame, either by calling tools, dynamically selecting the next action to take or use a deterministic "workflow" (systems where LLMs and tools are orchestrated through predefined code paths). Using an Agent framework that gives you the ability to built systems in both, a reliable and a dynamic way can accelerate your agent development journey using agent abstractions. This means building Agentic systems that can reliably call tools, store memory (episodic/semantic/procedural), have comprehensive logging and observability, human in the loop workflows and the ability to build various multi-agent patterns flexibly based on your use case. A successful Agentic system in production is usually a combination of both, dynamic and predictable/reliable multi-agent systems. Strands Agents SDK gives you exactly these capabilities by treating each “agent” as a combination of a foundation model plus a suite of tools. You define a prompt and register your tools (decorated functions) in code, then Strands handles reasoning→planning→tool-execution cycles, local testing, and cloud deployment (ECS, Fargate, Lambda, EC2), along with support for all other agent abstractions provided above. Excited to share a hedge-fund analyst multi agent system: This uses the newest Anthropic's Claude 4 Sonnet/Opus that powers the Lead Analyst Agent, routing incoming queries to specialized sub-agents for fundamental, technical, and market analyses. Each specialist is wrapped as a callable tool (using the “agents-as-tools” multi-agent pattern), so the orchestrator never has to implement domain logic itself and can handoff the task to an agent as a tool. For sensitive operations (insider lookups), we utilize a HITL approval step that halts execution until a human grants consent. We also use meta-tooling that enables the Lead Analyst to generate, load, and invoke new custom tools at runtime—whether it’s a portfolio beta calculator or a pricer—without redeployment. Strands also embeds observability (Langfuse) and OpenTelemetry tracing so you can trace reasoning events, tool invocations, errors, and end-to-end workflows in real time. View more information on the code implementation here: https://lnkd.in/gJmwVyGi Code implementation: https://lnkd.in/gzTtJvJq Thanks to 🏄‍♂️ Cagatay Cali for being a reviewer/collaborator on this! Feel free to try it out and reach out with any questions/ideas. #aws #agenticAI #strands #agents #generativeAI

60

6 Comments

Today we announced new capabilities for our Shop Direct and Buy for Me experiences on Amazon, enabling now feeds to facilitate external merchants to list reliably their product listings and reach millions of Amazon customers. Excited on this feature, that is a win/win experience for customers, merchants and brands. Excited to be part of such important project.

27

Optimizing OpenSearch: Shard Strategy and Best Practices Proper shard strategy is crucial for constructing high-performance, scalable, and dependable OpenSearch clusters. AWS OpenSearch Service offers guidance to assist architects in designing for efficiency and resilience. Key Takeaways from AWS Best Practices: - Right-Size Your Shards: Aim for 20–50 GB per shard initially to balance query performance, memory usage, and cluster management overhead. - Primary vs Replica Shards: Total shards = Primary shards × (1 + number of replicas). Replicas enhance read performance and fault tolerance. - Shard Count Considerations: Too many small shards lead to higher cluster overhead and slower recovery. Too few large shards result in slower queries and limited parallelism. - Plan for Growth: Anticipate future data volume, query patterns, and indexing frequency. Avoid default "one-size-fits-all" settings; customize based on workload. - Distribution Across Nodes: Distribute shards evenly across nodes to avoid hot spots. Monitor shard allocation and rebalance as needed. - Reindexing Strategy: Over time, adjust shard count or size through reindexing to uphold optimal cluster performance. Why This Matters: A well-considered sharding strategy ensures your OpenSearch cluster can manage high-volume, low-latency queries, scale effectively, and maintain resilience even during node failures. Reference: AWS OpenSearch Service – Shard Strategy Best Practices https://lnkd.in/gsrM6RSs

12

Closing the week by celebrating another great customer story — this time from Kakao Corp In this latest blog, Kakao shares how they overcame infrastructure scaling bottlenecks by moving to JAX + TPUs, enabling faster training of their Kanana model family. Their goal was to build more efficient, adaptable AI systems at production scale. With the right hardware + software combination — JAX for flexibility and TPUs for performance — customers can unlock breakthrough speed, efficiency, and scale in their AI workloads. Excited to help bring these innovations to organizations worldwide. https://lnkd.in/gu8UAk57 #AIleadership #Jax #TPUs #ScalableAI

10