Weifeng Cui

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

🚀 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

𝗧𝗵𝗲 𝗯𝗲𝘀𝘁 𝗔𝗜 𝗰𝗼𝗱𝗶𝗻𝗴 𝗹𝗲𝘀𝘀𝗼𝗻 𝗰𝗼𝘀𝘁 𝗼𝘂𝗿 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿 $𝟲𝟬𝟬 𝗮𝗻𝗱 𝗮 𝗺𝗮𝗿𝗿𝗶𝗮𝗴𝗲 𝗮𝗿𝗴𝘂𝗺𝗲𝗻𝘁. Our VP of Engineering, Xiaofan(James) Luan, was supposed to buy his wife a Dior bag for their anniversary.  Instead, he bought three Claude Code subscriptions and spent the holiday trying to cross-compile 2 million lines of C++. Every fix on one platform broke two others.  $600 later, the only output was "git reset --hard" — and a very cold dinner table.😂 "Make it compile on Windows" is a trap. The real goal was "compile everywhere without hacks" — no AI is going to figure that out for you at 2 am. What worked: constraints before code, review tests not code, bottom-up, one layer at a time. Same task, two days. Then he ran six parallel Claude sessions across three machines with git worktree. The bottleneck stopped being intelligence and started being how fast one person can alt-tab. AI solves exactly the problem you give it. Engineering is in knowing which one to give. His wife is still waiting for that bag. Full story: https://lnkd.in/gtsW_Wvk ——— Follow Milvus, created by Zilliz, for everything related to unstructured data

9

Excited to share our team's work on developing a Ranking Engineer Agent (REA), an autonomous AI agent, to fundamentally transform the end-to-end ML lifecycle for ads ranking, one of the most sophisticated and complex machine learning systems. Improving ads ranking has been a deeply complex, manual, and multi-week-long process, heavily reliant on manual effort across hypothesis formation, configuration, execution, and failure recovery. REA is changing that paradigm. It now autonomously manages the entire workflow from generating high-impact hypotheses and designing experiments to executing training jobs, diagnosing system failures, and optimizing iteratively. This critical shift allows our engineers to pivot from operating the experiment loop to directing it: setting the strategic vision, evaluating macro-level results, and making crucial architectural decisions that propel our business forward. Congratulations to the entire team for pushing the boundaries of what autonomous ML engineering means at Meta scale. Please see the full technical deep dive on the Meta Engineering Blog here. https://lnkd.in/gtjGyrBv

265

1 Comment

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

this is super important because TPUs have definitely advantages for cost and performance over other accelators but were much more difficult to use. no longer the case. "vLLM TPU is now powered by tpu-inference, an expressive and powerful new hardware plugin unifying JAX and PyTorch under a single lowering path. It is not only faster than the previous generation of vLLM TPU, but also offers broader model coverage and feature support." https://lnkd.in/diDKzYbS

35

1 Comment

Tuning Spark, Ray, or Dask for ML isn’t one-size-fits-all. In a new Domino Blueprint, Principal Solution Architect Sameer Wadkar shows how API-driven scaling complements autoscaling to right-size clusters on demand—cutting idle spend and speeding bursts for GPU/I/O-heavy work. Practical steps, a lightweight Python client, and Helm install included. If you care about FinOps and MLOps, this is for you. Read the blueprint: https://lnkd.in/ga_qJkhC

7

Apache Spark’s new "Real-time" Mode in Structured Streaming delivers millisecond-level latency for ultra-fast, continuous stream processing. It will be interesting to see how it will perform compared to Flink!

7