---
title: "Latest reads"
description: "The engineering internet, summarised so you can actually read it."
---

# Latest reads
> The engineering internet, summarised so you can actually read it.

## Articles

### [Bringing custom knowledge to agents with AutoRAG](https://yomu.fyi/post/bringing-custom-knowledge-to-agents-with-autorag.md)
- Company: [Red Hat](https://yomu.fyi/company/red-hat.md)
- Author: Diego Alvarez Ponce, Diego Garcia Perez
- Published: Sep 11, 2026
- Type: Tutorial

Enterprise LLMs can hallucinate when they lack company-specific vocabulary and documentation, so the guide builds a bank agent grounded in internal data. It configures Red Hat OpenShift AI 3.5 Technology Preview features, enables OGX, AutoRAG, and the MCP catalog, and deploys PostgreSQL, Milvus, and MinIO for metadata, vector embeddings, and artifacts. AutoRAG processes documents, evaluates pipeline configurations against a benchmark dataset, and identifies a preferred setup; the example uses Qwen3-8B and embeddinggemma-300m models. A MariaDB MCP server adds access to customer records, with a patch supplying writable logging storage and a warning that unrestricted ALLOWED\_HOSTS is for demo/development only. The resulting Streamlit agent answers a customer-information query and a Gold-card-limit question using RAG context and database tools.


### [Red Hat edge platforms: Choosing the right one for your use case](https://yomu.fyi/post/red-hat-edge-platforms-choosing-the-right-one-for-your-use-case.md)
- Company: [Red Hat](https://yomu.fyi/company/red-hat.md)
- Author: Daniel Froehlich
- Published: Sep 11, 2026
- Type: Explainer

Choosing an edge platform requires balancing hardware footprint, workload needs, high availability, and operational complexity; the options range from RHEL with Podman or KVM to OpenShift clusters. The guide organizes Red Hat's offerings into RHEL-based and OpenShift-based deployments, noting that both traditional virtual machines and modern containerized workloads can coexist and that platforms can overlap as requirements grow. It details image-mode RHEL (bootc), transactional updates with rollbacks, efficient over-the-air delivery, offline installation, and dependency pre-resolution for constrained or disconnected sites. RHEL is positioned for static workloads of one to five components, while Red Hat Device Edge adds MicroShift when lightweight Kubernetes is needed. Kubernetes becomes more appropriate for larger microservices or event-driven systems, automated recovery, dynamic workloads, cross-environment scaling, and rolling updates; the series next covers OpenShift topologies.


### [Introducing automatic remediation policies with Cloudflare CASB](https://yomu.fyi/post/introducing-automatic-remediation-policies-with-cloudflare-casb.md)
- Company: [Cloudflare](https://yomu.fyi/company/cloudflare.md)
- Author: Michael Leslie
- Published: Sep 11, 2026
- Type: Announcement

Cloudflare CASB now supports automatic remediation policies that respond to newly detected SaaS security findings without requiring an administrator to act on each instance. Teams can match policies by vendor, integration or tenant, and finding type, then revoke file-sharing access, send finding details to webhooks, or do both; first-party remediation currently covers Microsoft and Google Workspace file and folder findings. The backend enqueues findings to Cloudflare Queues, uses a Worker to match policy configurations, and runs matched jobs through Cloudflare Workflows, which provides durable execution, retries, and rate-limit backoff. Policies generate administrator-activity and runtime execution logs, while the stated target from detection to completed remediation is five minutes or less; support for Custom Findings is planned for the coming weeks.


### [AI Functions in ClickHouse: Upgrade your SQL to the AI age](https://yomu.fyi/post/ai-functions-in-clickhouse-upgrade-your-sql-to-the-ai-age.md)
- Company: [Clickhouse](https://yomu.fyi/company/clickhouse.md)
- Published: Sep 11, 2026
- Type: Explainer

ClickHouse’s beta AI Functions let SQL queries call an LLM or embedding provider directly, keeping classification, extraction, generation, translation, redaction, filtering, embedding, and similarity work alongside stored data. The feature addresses workflows that export rows to separate LLM, orchestration, or vector-database systems and then reload results, with the stated goal of running the full RAG cycle in ClickHouse. Functions make remote HTTP calls per input, or batch embedding inputs, and return native ClickHouse values; in open-source deployments, named collections store provider credentials and endpoint settings. Examples use Hacker News data for topic classification and natural-language filtering, while the guidance emphasizes cheap predicates, quotas, LIMIT tests, prompt-injection defenses, non-determinism, cost, latency, and provider access to data in cleartext after TLS termination.


### [Debugging our AI search assistant with agent tracing](https://yomu.fyi/post/debugging-our-ai-search-assistant-with-agent-tracing.md)
- Company: [Sentry](https://yomu.fyi/company/sentry.md)
- Author: Dominik Buszowiecki, Shaun Kaasten
- Published: Sep 11, 2026
- Type: Explainer

Sentry’s Search Query Assistant converts natural-language prompts into Sentry Search Syntax, but its non-deterministic output requires evals and debugging when generated queries are wrong. A production issue caused searches on custom numerical attributes to return no events, so the team used Sentry’s AI Conversation view and Agents filters to find the relevant trace, inspect the timeline, and follow tool calls into the backend API. The trace showed that tool results were being incorporated into the system prompt with the field name bug\_\_predictions\_count instead of the required tags\[bug\_predictions\_count, number\] form. The team added an eval scenario, fixed the prompt generation, verified the corrected query locally and in production, and confirmed the updated prompt and API parameter in the trace. They also instrumented local eval runs to send data to Sentry, allowing debugging before a pull request.


### [Together AI expands fine-tuning service with more models, live metrics, and finer controls](https://yomu.fyi/post/together-ai-expands-fine-tuning-service-with-more-models-live-metrics.md)
- Company: [Together AI](https://yomu.fyi/company/together-ai.md)
- Published: Sep 11, 2026
- Type: Announcement

Together AI is expanding its Fine-Tuning service to cover newer open-weight models and more of the workflow from dataset inspection through model selection. The update adds live training and evaluation metrics through the API, CLI, dashboard, and Python SDK, including curves that can be compared across in-flight jobs and underlying series that can be exported. Training controls now include Expert LoRA for mixture-of-experts layers, early stopping based on validation loss, and arbitrary effective batch sizes through gradient accumulation. The post reports that Expert LoRA recalled up to 89% of 200 invented facts versus 15% for attention-only adapters, while listed LoRA training prices fall by 30% to 70% for selected models, including the gpt-oss series.


### [Welcoming Julia Zettel, our new Video Producer](https://yomu.fyi/post/welcoming-julia-zettel-our-new-video-producer.md)
- Company: [Resend](https://yomu.fyi/company/resend.md)
- Published: Sep 11, 2026
- Type: Announcement

Resend announces that Julia Zettel has joined its team as a Video Producer, bringing more than a decade of experience in video and storytelling from work with brands including Zee.Dog and Globo. She will help the company communicate with customers, produce original content, and tell stories beyond the product through video, education, launches, and the people building with Resend. Julia says her background grew from being surrounded by television and music production, while her studies in Social Communication and AV led her toward narrative-driven visual work; she initially collaborated with Resend on editing, then a full producing and filming project before joining full time. She cites Adobe Premiere as her primary tool, Adobe Podcast for noisy or windy recordings, and recommends accepting challenging projects and learning from experienced people alongside online tutorials and AI.


### [How to Connect Your Twilio Agent to External APIs with PHP](https://yomu.fyi/post/how-to-connect-your-twilio-agent-to-external-apis-with-php.md)
- Company: [Twilio](https://yomu.fyi/company/twilio.md)
- Author: Amanda Lange, Matthew Setter
- Published: Sep 11, 2026
- Type: Tutorial

The tutorial shows how to connect a Twilio voice agent to external REST APIs so an LLM-driven conversation can retrieve live information or support actions such as checking inventory or booking appointments. It builds the example in PHP 8.5 or later with OpenSwoole, Twilio Conversation Relay, OpenAI, and a WebSocket server, using Cat Facts as an unauthenticated API. The application exposes a POST /voice webhook that returns TwiML and connects Conversation Relay to a wss /ws endpoint, while session handling passes caller prompts into the OpenAI service. Tool calling routes the request to the Cat Facts endpoint, parses the JSON fact field, and returns either the fact, “No fact returned.”, or an error when the API response is unsuccessful or unavailable. Testing requires running php server.php, exposing port 8080 with ngrok, and configuring the Twilio number’s incoming webhook to /voice; the conclusion says external HTTP endpoints give agents real-time data.


### [How to Orchestrate Multi-Call Conversations with an LLM and Twilio Conversation Memory with PHP](https://yomu.fyi/post/how-to-orchestrate-multi-call-conversations-with-an-llm-and-twilio-con.md)
- Company: [Twilio](https://yomu.fyi/company/twilio.md)
- Author: Amanda Lange, Matthew Setter
- Published: Sep 11, 2026
- Type: Tutorial

This tutorial shows how to build a PHP voice agent that retains caller context, preferences, and action history across separate inbound calls. It uses Twilio Conversation Memory with a Memory Store, an Open Swoole service, Twilio ConversationRelay, and an OpenAI integration using gpt-4o-mini; the example models an auto-repair shop assistant. The OpenAI service appends supplied prior customer context to its system prompt, while streamed responses are sent through a WebSocket and environment variables provide API credentials and the Memory Store ID. After running the service on port 9501, exposing it with ngrok, and configuring the phone-number webhook, the tutorial tests two calls from the same number and says the second call should remember details from the first.


### [Adding Google Login to a React App with Auth0](https://yomu.fyi/post/adding-google-login-to-a-react-app-with-auth0.md)
- Company: [Auth0](https://yomu.fyi/company/auth0.md)
- Author: Carla Urrea Stabile
- Published: Sep 11, 2026
- Type: Tutorial

The tutorial shows how to add Google login to a Vite-based React single-page app without implementing OAuth directly, using Auth0’s React SDK and Google social connection. It covers installing @auth0/auth0-react, creating an Auth0 Single Page Web Application, checking RS256 and OIDC Conformant settings, registering callback, logout, and web-origin URLs, and supplying the domain and client ID through Vite environment variables. The React code wraps the app in Auth0Provider, uses useAuth0 for loading and authentication state, and calls loginWithRedirect with connection: 'google-oauth2' to bypass provider selection. For production, it explains replacing Auth0’s shared development keys with Google Cloud credentials whose origin and redirect URI target the Auth0 domain, then storing them in the Auth0 social connection. It also points to route protection with withAuthenticationRequired, API access tokens from getAccessTokenSilently, and exact redirect-URI matching as important follow-on considerations.


### [DiscoSign: Discourse-Aware Text to Sign Language Gloss Translation](https://yomu.fyi/post/discosign-discourse-aware-text-to-sign-language-gloss-translation.md)
- Company: [Apple](https://yomu.fyi/company/apple.md)
- Published: Sep 11, 2026
- Type: Problem & solution

DiscoSign addresses a limitation in sign language processing: systems have traditionally translated at the sentence level, overlooking discourse phenomena important to sign language comprehension. It presents a modular Large Language Model (LLM)-based framework for text-to-sign language gloss translation that handles spatial coreference resolution, Question-Answer Clauses (QACs), and concept-gloss consistency, preserving entity locations and stable English-to-ASL sign mappings across discourse. Because traditional translation metrics do not measure discourse-level quality, the authors introduce metrics for each coherence dimension targeted by the framework. Experiments on sentence-level and discourse-level datasets report improved spatial consistency and entity tracking over sentence-only translation, while maintaining competitive single-sentence gloss quality; the work describes this as the first systematic framework with a corresponding evaluation methodology.


### [Putting Captions to the Test: Evaluating Video Caption Quality through Multiple-Choice Question Answering](https://yomu.fyi/post/putting-captions-to-the-test-evaluating-video-caption-quality-through.md)
- Company: [Apple](https://yomu.fyi/company/apple.md)
- Published: Sep 11, 2026
- Type: Benchmark

Evaluating video captions is difficult because reference-matching metrics can penalize valid descriptions that use different wording or focus on different visual content, while offering little fine-grained diagnosis. The paper reframes quality around information fidelity: captions should cover salient visual information and remain factually accurate. It introduces CapQuiz, a reference-free benchmark that tests caption utility through human-verified, fine-grained multiple-choice questions derived from videos, organized into 10 Descriptive and Inferential question types across 24 domains. It also proposes CapF1, which combines CapP for factuality with CapR for coverage; extensive experiments report significantly stronger correlation with human judgments than existing metrics and interpretable insight into model performance.


### [SimpleDesign: A Joint Model for Protein Sequence and Structure Codesign](https://yomu.fyi/post/simpledesign-a-joint-model-for-protein-sequence-and-structure-codesign.md)
- Company: [Apple](https://yomu.fyi/company/apple.md)
- Published: Sep 11, 2026
- Type: Problem & solution

SimpleDesign addresses protein design as a joint modeling problem involving amino acid sequences and three-dimensional structures, whose interplay determines protein function. Unlike approaches that first train autoencoders and then model their latent representations, it uses a single-stage, end-to-end objective directly in data space. The objective combines discrete cross-entropy for sequences with a regression objective for structures, while Transformer-based multimodal backbones provide modality-specific processing and global self-attention across both modalities. Trained on more than 2M sequence-structure pairs, SimpleDesign achieves competitive performance on co-design and unconditional sequence and structure generation benchmarks. The work argues, based on these results, that multi-stage training is not necessary for performant co-design models.


### [Autonomous LLM post-training with Tunix on TPUs](https://yomu.fyi/post/autonomous-llm-post-training-with-tunix-on-tpus.md)
- Company: [Google](https://yomu.fyi/company/google.md)
- Author: Wei Wei
- Published: Sep 11, 2026
- Type: Benchmark

autofinetune applies autonomous research loops to LLM post-training, covering supervised fine-tuning (SFT) and reinforcement learning with GRPO through Tunix, Gemma, Cloud TPUs, Antigravity CLI, and Gemini Flash 3.7. A human defines the arena in program.md and supplies a self-contained run.py, while an agent edits hyperparameters, launches jobs, evaluates a target metric, keeps winning Git commits, reverts regressions, and logs results in results.tsv. On FunctionGemma 270M with google/mobile-actions, 20 Cloud TPU v5e-1 experiments tested LoRA settings, optimization choices, schedules, batch size, and seeds without changing the dataset, epochs, or architecture. A second study used Gemma 3 1B, GSM8K, and Cloud TPU v6e-1 for 40 GRPO experiments over two to three days, improving the combined numerical-and-format reward by about 10%.


### [A Developer's Guide to API Access Policies in Auth0](https://yomu.fyi/post/a-developer-s-guide-to-api-access-policies-in-auth0.md)
- Company: [Auth0](https://yomu.fyi/company/auth0.md)
- Author: Madhuri Ravindra Mohan
- Published: Sep 11, 2026
- Type: Tutorial

Auth0 API Access Policies for Applications control whether an application can obtain an access token for a registered API, separately for user-delegated and client access. The FinFlow example separates a browser application that should reach both APIs from a machine-to-machine Risk Engine restricted to the Payments API. Policies provide three choices—allow\_all, require\_client\_grant, and deny\_all—while client grants specify an application's audience and permitted scopes, with the audience determining which API policy is evaluated. The guide shows Dashboard and Management API configuration, including subject\_type\_authorization settings and grants for user and client access. It also explains that third-party applications still require grants under allow\_all, while default third-party grants establish a baseline and per-application grants take precedence for exceptions.


### [\[Webinar\] Elevating customer engagement with WhatsApp](https://yomu.fyi/post/webinar-elevating-customer-engagement-with-whatsapp.md)
- Company: [Twilio](https://yomu.fyi/company/twilio.md)
- Author: Twilio
- Published: Sep 11, 2026
- Type: Announcement

The webinar examines how businesses can use WhatsApp for secure account verification, one-time passcodes, operational alerts, and customer conversations. It presents notification overload, third-party portals, single-use apps, and unbranded sender numbers as reasons critical messages may be ignored, affecting deliveries and transactions. Speakers from Twilio, WhatsApp, and Orbio AI describe a verified, two-way model built around WhatsApp Utility messaging, interactive buttons, logic flows, end-to-end encryption, and visual drag-and-drop automation. Orbio AI's examples include screening candidates, scheduling interviews, reminders, onboarding, and engagement conversations for enterprise clients while the webinar claims improved deliverability, faster user action, reduced manual work, protection against SMS toll fraud, and scalable deployment. It also positions app-based delivery over WiFi and cellular networks as a more reliable alternative.


### [Building Pinterest’s VLM Serving Stack on NVIDIA Dynamo](https://yomu.fyi/post/building-pinterest-s-vlm-serving-stack-on-nvidia-dynamo.md)
- Company: [Pinterest](https://yomu.fyi/company/pinterest.md)
- Author: Pinterest Engineering
- Published: Sep 10, 2026
- Type: Explainer

Pinterest describes building a VLM serving stack on NVIDIA Blackwell B200 GPUs and NVIDIA Dynamo for multimodal products including Pinterest Assistant, hybrid search, multimodal reranking, content understanding, signal generation, and content safety. Unlike text-only workloads, requests may contain thousands of images, costly and variable prefill, long multi-turn histories, and large KV caches, making preprocessing, routing, memory, and latency central serving concerns. The implementation combines Dynamo's distributed orchestration, multimodal frontend and vLLM support, KV-aware routing, disaggregated encoder/prefill/decode serving, and LMCache-based KV offloading with vLLM, an OpenAI Chat Completions interface, model-aware routing, EKS, and custom deployment components. Production-oriented AIPerf DAG benchmarks model QPS patterns, multi-turn conversations, many-image requests, and agentic DAGs; the team concludes that payload design, cache-aware routing, offload tiers, and a shared platform are essential to meeting product targets and accelerating new launches.


### [Halving the Time: How Uber Eats Rebuilt Its Search Pipeline](https://yomu.fyi/post/halving-the-time-how-uber-eats-rebuilt-its-search-pipeline.md)
- Company: [Uber](https://yomu.fyi/company/uber.md)
- Published: Sep 10, 2026
- Type: Problem & solution

Uber Eats rebuilt its search pipeline after accumulated architectural decisions made latency increasingly costly across retrieval, hydration, ranking, ads, and presentation. The team changed its primary metric from backend API response time to Above-the-Fold completion, then introduced paginated responses with server-side caching and concurrent template rendering; these changes improved ATF by over 200 milliseconds. Other work removed low-yield retrieval strategies, grouped items by product identity to reduce lookups by over 100 times, split ranking and presentation hydration, eliminated false DAG dependencies, added request hedging, and redesigned ad data as column-oriented in memory. The combined effort cut search latency in half, including approximately 120 milliseconds from retrieval, 100-plus from hydration changes, and 130 from ads, while planned product-based search, zero-pass ranking, and streaming target further gains.


### [Grocery Delivery across the US: Fees, Pricing and Membership Guide](https://yomu.fyi/post/grocery-delivery-across-the-us-fees-pricing-and-membership-guide.md)
- Company: [Uber](https://yomu.fyi/company/uber.md)
- Published: Sep 10, 2026
- Type: Explainer

Uber Eats grocery orders combine item prices with Delivery, Service, and possible add-on fees, while taxes are estimated at checkout and finalized on the receipt. Delivery Fees typically range from $0 with Uber One on eligible orders to $7.99 for standard orders, and Service Fees typically range from $2.85 to $7.75 depending partly on basket size. Customers can order without a membership; participating stores may offer in-store pricing, while other stores may mark up item prices. Uber One costs $9.99 monthly or $96 annually, provides $0 Delivery Fees on eligible grocery orders, and is described as breaking even after approximately two grocery orders per month. It also extends benefits to restaurant delivery and other Uber services, while listed savings options include promotions, No Rush Delivery, larger consolidated orders, EBT payments, and discounted gift cards.


### [McDonald's x Uber Eats Courier Appreciation Week Rules](https://yomu.fyi/post/mcdonald-s-x-uber-eats-courier-appreciation-week-rules.md)
- Company: [Uber](https://yomu.fyi/company/uber.md)
- Published: Sep 10, 2026
- Type: Announcement

The McDonald’s Courier Appreciation Week Sweepstakes is sponsored and administered by Uber and is limited to eligible U.S. Uber Eats delivery drivers. Participants must be at least 18, reside in Uber serviceable areas, hold a delivery account as of September 14, 2026, meet delivery-driver requirements, and have an account in good standing. During the September 14–18 Sweepstakes Period, invited drivers can receive one entry by completing the McDonald’s Courier Feedback Survey, with 100 potential winners selected randomly on or about September 25. Each winner receives $50 in Uber Cash, for a total approximate prize value of $5,000, subject to usage restrictions and possible expiration; no purchase is necessary, and odds depend on the number of eligible entries.


[Newer posts](https://yomu.fyi/posts/page/46.md) · [Older posts](https://yomu.fyi/posts/page/48.md)