--- title: Netflix maintains an LLM judge in production, Uber tames... description: Today's posts skew toward the unglamorous work of running AI systems at scale rather than announcing new ones. Netflix published how it keeps an LLM judge from... canonical: https://daily.dev/posts/netflix-maintains-an-llm-judge-in-production-uber-tames-a-7x-agent-cost-surge-ihigsty2t twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Netflix maintains an LLM judge in production, Uber tames a 7x agent cost surge | daily.dev og:description: Today's posts skew toward the unglamorous work of running AI systems at scale rather than announcing new ones. Netflix published how it keeps an LLM judge from... og:url: https://daily.dev/posts/netflix-maintains-an-llm-judge-in-production-uber-tames-a-7x-agent-cost-surge-ihigsty2t og:image: https://api.daily.dev/og/posts/IHIGsTY2t.png og:image:alt: Netflix maintains an LLM judge in production, Uber tames a 7x agent cost surge og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Netflix maintains an LLM judge in production, Uber tames a 7x agent cost surge **[Backend Digest](https://daily.dev/sources/backend_digest)** · 4 min read · 1 upvotes · 0 comments ## Summary Today's posts skew toward the unglamorous work of running AI systems at scale rather than announcing new ones. Netflix published how it keeps an LLM judge from drifting out of alignment months after launch. Uber shared the cost-accounting playbook that kept agentic coding spend flat while usage grew 7x. Atlassian and Cloudflare both offered concrete infrastructure writeups worth bookmarking, and a couple of RAG and database posts round things out. ## Content **TLDR:** Today's posts skew toward the unglamorous work of running AI systems at scale rather than announcing new ones. Netflix published how it keeps an LLM judge from drifting out of alignment months after launch. Uber shared the cost-accounting playbook that kept agentic coding spend flat while usage grew 7x. Atlassian and Cloudflare both offered concrete infrastructure writeups worth bookmarking, and a couple of RAG and database posts round things out. --- ## Netflix details how it keeps an LLM-as-judge from drifting in production Netflix uses one model to write "because you watched" blurbs and a second model to grade every one before it ships. The new paper breaks judge maintenance into four stages, and the interesting bit is the last one: weekly drift monitoring calibrated against how much human raters disagree with each other, not some fixed accuracy target. They also train the judge on written rationales instead of pass/fail labels, so it learns to reject bad output for the right reason. In a 5-week test, adding explanations back in produced a small but real shift toward previously unwatched titles. If you've got a judge model gating anything real, this is the maintenance discipline most teams skip. [Read more](https://daily.dev/posts/J7XUSD54H) ## Uber holds agentic coding costs flat through a 7x usage spike Weekly active users on Uber's agentic coding tools grew 7x and requests grew 9.4x since February, yet costs stayed stable. The trick was decomposing session cost into users, sessions, turns, tokens, and price-per-token, then squeezing each one separately: cheaper models for subagents, 400K token context caps, CLI-resolved MCP tool calls instead of preloading schemas (saving 50-70K tokens per call), and code-mode batching that cuts token usage over 90% for bulk workflows. A session dashboard now flags 16 categories of wasted spend, including bad model routing and cache expiration. [Read more](https://daily.dev/posts/fyfellEFf) ## Atlassian's agent pipeline merges 55 vulnerability-fix PRs with a 95% first-try success rate Atlassian combined Bitbucket Agentic Pipelines, Jira Automation, and Rovo Dev to automate routine security patching: a dispatcher groups eligible tickets, a coding agent opens PRs, and a closer verifies deployment before closing the ticket. Over three months one team resolved 120+ vulnerabilities and merged 55+ agent-created PRs, with 95% merging clean on the first run. The scoped-access and deterministic-verification setup is the part worth stealing regardless of which agent runner you use. [Read more](https://daily.dev/posts/XXtgsat4y) ## Cloudflare explains why it bet on eBPF over kernel bypass for its entire networking stack Cloudflare's engineering team recounts abandoning a proprietary RDMA kernel-bypass stack after it locked them into one NIC vendor and broke during an ARM/25G upgrade. XDP delivered vendor-agnostic packet dropping past 10M pps at roughly 10% CPU overhead during attacks, while keeping standard Linux tooling intact, unlike DPDK-style full bypass. Six production systems now run on eBPF, covering DDoS mitigation, load balancing, socket dispatch, firewalling, observability, and reliability — a useful reference if you're weighing kernel bypass against a programmable-networking approach. [Read more](https://daily.dev/posts/qCDggFpuW) --- ## Also notable - **Switching LLM providers breaks on 12 specific gaps:** A practical catalog covers inconsistent max_tokens handling, missing streaming usage data, and non-interchangeable embeddings across OpenAI-compatible APIs, with a probe script to test a provider before depending on it. [Read more](https://daily.dev/posts/pJoQ7dpC1) - **SQLite FTS5 + RRF replaces a vector database and reportedly gets 10x better RAG results:** A ~50-line hybrid search implementation fuses BM25 lexical search with dense embeddings at RRF k=60, with the author arguing dedicated vector databases like Milvus or Pinecone only earn their keep past roughly 1 million chunks. [Read more](https://daily.dev/posts/ChwXB9NhH) - **GPU-pruner scales idle Kubernetes GPU workloads to zero using DCGM metrics:** The open-source controller monitors actual GPU engine utilization with a 35-minute idle window and 1% activity threshold, then walks ownerReferences to pause the right Deployment or InferenceService instead of deleting it. [Read more](https://daily.dev/posts/LLycyb0Ay) - **Rails uniqueness validations can't stop race conditions without a DB-level unique index:** Only a database unique index actually arbitrates concurrent writes, since the validator's SELECT and the following INSERT are separate steps another connection can interleave with, raising ActiveRecord::RecordNotUnique on the losing insert. [Read more](https://daily.dev/posts/fSZaWH3dQ) - **node:sqlite reaches release-candidate status in Node 24.15:** The built-in SQLite module needs no native compilation or npm dependency, but its fully synchronous execution model can block the JS thread on busy HTTP servers, so treat newer parts of the API with caution in long-lived services. [Read more](https://daily.dev/posts/wIUjo9T6G) --- Tags: [#linux](https://daily.dev/tags/linux), [#ai-agents](https://daily.dev/tags/ai-agents), [#rag](https://daily.dev/tags/rag), [#llmops](https://daily.dev/tags/llmops) [View this post on daily.dev](https://daily.dev/posts/netflix-maintains-an-llm-judge-in-production-uber-tames-a-7x-agent-cost-surge-ihigsty2t) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"DiscussionForumPosting","mainEntityOfPage":"https://daily.dev/posts/netflix-maintains-an-llm-judge-in-production-uber-tames-a-7x-agent-cost-surge-ihigsty2t","headline":"Netflix maintains an LLM judge in production, Uber tames a 7x agent cost surge","text":"Today's posts skew toward the unglamorous work of running AI systems at scale rather than announcing new ones. Netflix published how it keeps an LLM judge from drifting out of alignment months after launch. Uber shared the cost-accounting playbook that kept agentic coding spend flat while usage grew 7x. Atlassian and Cloudflare both offered concrete infrastructure writeups worth bookmarking, and a couple of RAG and database posts round things out.","url":"https://daily.dev/posts/netflix-maintains-an-llm-judge-in-production-uber-tames-a-7x-agent-cost-surge-ihigsty2t","datePublished":"2026-08-31T04:18:58.037Z","dateModified":"2026-08-31T04:19:20.771Z","author":{"@type":"Organization","name":"Backend Digest","logo":"https://media.daily.dev/image/upload/s--9uCSrsAb--/f_auto,q_auto/v1773839398/logos/backend_digest?_a=BAMAMiiu0","url":"https://daily.dev/sources/backend_digest"},"interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"isPartOf":{"@type":"WebPage","url":"https://daily.dev/sources/backend_digest","name":"Backend Digest"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Backend Digest","item":"https://daily.dev/sources/backend_digest"},{"@type":"ListItem","position":3,"name":"Netflix maintains an LLM judge in production, Uber tames a 7x agent cost surge"}]} ```