---
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

### [How AI Agents Get Trusted Customer Context with Data 360 Data Graphs](https://yomu.fyi/post/how-ai-agents-get-trusted-customer-context-with-data-360-data-graphs.md)
- Company: [Salesforce](https://yomu.fyi/company/salesforce.md)
- Author: Scott Nyberg
- Published: Sep 14, 2026
- Type: Explainer

Salesforce’s AI Engineering team uses Data 360 data graphs to close the context gap that prevents agents from reliably personalizing customer interactions. The graphs join, aggregate, and apply business logic across fragmented identities, accounts, products, contracts, orders, cases, services, and entitlements, while partitioned data spaces provide a filtered customer-success view. The team designs smaller graphs or multi-graphs around agent access patterns, adds indices, supports semantic or keyword search, and curates context so agents can traverse extensible many-to-many relationships without receiving unnecessary data. Reusable deployment automation and validation practices helped deliver five major releases in six months, with integrations spanning Postman, Apex, Agentforce, and low-code or no-code workflows. Live monitoring later showed P50 performance below 200 milliseconds, compared with an earlier benchmark of about 400 milliseconds.


### [Accelerating Dropless MoE Training in JAX with NVIDIA Transformer Engine](https://yomu.fyi/post/accelerating-dropless-moe-training-in-jax-with-nvidia-transformer-engi.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Tanya Lenz
- Published: Sep 14, 2026
- Type: Explainer

Production-scale mixture-of-experts (MoE) training must handle uneven, data-dependent expert loads, creating ragged tensors and expensive dispatch, combine, and all-to-all communication. Dropless MoE processes every selected token, preserving model quality but requiring kernels that support variable token counts without padding or dropping. The post presents JAX and NVIDIA Transformer Engine optimizations centered on group-aware MXFP8 quantization, grouped GEMM, and fused expert-parallel (EP) dispatch and combine using NCCL EP. Grouped GEMM computes each expert’s actual workload in one kernel path, while EP reduces movement overhead and can deduplicate tokens crossing networks. For DeepSeek-V3 training on NVIDIA GB200, the reported throughput increased from 103 to 1,068 TFLOPS/GPU, a 10.4x improvement; the post also reports roughly 10x throughput improvement and 97% scaling efficiency at 1,024 GPUs.


### [Toolbox App 3.8: Improves IDE Update Handling on macOS and Fixes Keyboard Navigation](https://yomu.fyi/post/toolbox-app-3-8-improves-ide-update-handling-on-macos-and-fixes-keyboa.md)
- Company: [Jetbrains](https://yomu.fyi/company/jetbrains.md)
- Author: Ivan Kuzmin
- Published: Sep 14, 2026
- Type: Announcement

Toolbox App 3.8 addresses IDE update handling on macOS and restores keyboard navigation in project and tool lists. On macOS, manual IDE updates remain available when the system incorrectly reports an installation directory as non-writable, with instructions shown if macOS blocks the update. On Windows, Ctrl+1 and Ctrl+2 switch between the Tools and Projects tabs, while the Up Arrow, Down Arrow, Page Up, Page Down, Home, and End keys navigate lists, including filtered results. Cleanup of leftover IDE directories now displays progress for each tool and the full operation, and disables delete controls until cleanup and rescanning finish, preventing duplicate cleanup starts.


### [Understanding W8A8 INT8 LLM quantization: Accuracy and performance results](https://yomu.fyi/post/understanding-w8a8-int8-llm-quantization-accuracy-and-performance-resu.md)
- Company: [Red Hat](https://yomu.fyi/company/red-hat.md)
- Author: Sana Fayyaz
- Published: Sep 14, 2026
- Type: Benchmark

The post evaluates whether W8A8 INT8 quantization can reduce the size of a Llama 3.1 8B Instruct model without sacrificing accuracy or inference performance. Using SmoothQuant to address activation outliers and GPTQ with error compensation for BF16 weights, the model was compressed from 14.9 GB to 8.0 GB after calibration on WikiText-2. Identical zero-shot lm-eval-harness tests across MMLU, ARC Easy, HellaSwag, and IFEval found MMLU, ARC Easy, and HellaSwag deltas smaller than their standard errors, while the IFEval change was described as likely within a similar noise range; the post concludes no measurable degradation. vLLM and GuideLLM tests on one NVIDIA L40S reported 29% more concurrent requests, 24% lower synchronous TTFT, and 44% faster token generation at maximum load, while ITL degraded more quickly under high concurrency because of dynamic activation quantization overhead.


### [Python 3.14 free-threaded build is now available in RHEL](https://yomu.fyi/post/python-3-14-free-threaded-build-is-now-available-in-rhel.md)
- Company: [Red Hat](https://yomu.fyi/company/red-hat.md)
- Author: Lumír Balhar
- Published: Sep 14, 2026
- Type: Announcement

Red Hat Enterprise Linux 9.8 and 10.2 now provide a free-threaded Python 3.14 build alongside the regular interpreter, enabling developers to test parallel CPU execution with threads. The build disables the GIL, allowing Python threads to execute on multiple CPU cores in one process, while built-in dict, list, and set operations retain internal locking for individual operations. Developers install the python3.14-freethreading package from the CodeReady Linux Builder repository; it provides the separate python3.14t executable and commands to verify free-threading support and GIL status. Existing code may expose race conditions, and unsupported C-extension dependencies can re-enable the GIL, so the source recommends testing, dependency checks, explicit synchronization, and workload measurement. The build carries a stated 5%-10% single-threaded performance overhead and mainly targets CPU-bound threaded workloads.


### [Stop Rubber-stamping AI: A Contestable AI Framework for Regulated Industries](https://yomu.fyi/post/stop-rubber-stamping-ai-a-contestable-ai-framework-for-regulated-indus.md)
- Company: [MongoDB](https://yomu.fyi/company/mongodb.md)
- Published: Sep 14, 2026
- Type: Explainer

AI deployed in regulated industries such as healthcare and manufacturing often lacks explainability and contestability, creating a barrier to adoption because responsible humans cannot meaningfully challenge AI outputs. The post presents contestable AI as an architectural pattern where multiple specialist agents debate recommendations, a qualified human reviews and overrides arguments, and the system re-scores in real time. In a clinical scenario, a patient analyzer agent retrieves similar cases, specialists argue for and against care options, and a general practitioner overrides as needed before a care plan is finalized. In manufacturing, a site reader agent ingests equipment sensor labels, two specialist agents debate classifications against a standard ontology, and an engineer overrides with context the model lacked, like a fixture offset. MongoDB’s document model stores the AI arguments, human overrides, confidence scores, and final decision in one retrievable document, supporting single-query decision lineage without joining data across systems.


### [Product updates: Sandbox Sidecars, new models, a refreshed dashboard, and more](https://yomu.fyi/post/product-updates-sandbox-sidecars-new-models-a-refreshed-dashboard-and.md)
- Company: [Modal](https://yomu.fyi/company/modal.md)
- Published: Sep 14, 2026
- Type: Announcement

Modal added day-zero support for Kimi K3, Qwen 3.8, GLM 5.3, and GLM 5.3 Flash on Auto Endpoints, with custom DFlash speculators boosting throughput to 460 tokens per second. New environment-level budgets give Team and Enterprise workspaces per-environment compute limits, while the pricing multiplier for broad region selection dropped from 1.5x to 1.15x. The dashboard got a refreshed navigation, color system, and beta light mode, and Private Environments now support a default No Access role. Sandbox Sidecars are in public alpha, running additional containers via low-latency internal networking, and VM Sandboxes are in public beta with a real Linux kernel for Docker and eBPF. Sandbox directory snapshots and a new filesystem API are now GA, and regional Proxies allow selecting a region for static outbound IPs, while SDK updates improved log APIs and autoscaler reporting.


### [How to Orchestrate Multi-Call Conversations with an LLM and Twilio Conversation Memory in Python](https://yomu.fyi/post/how-to-orchestrate-multi-call-conversations-with-an-llm-and-twilio-con-python.md)
- Company: [Twilio](https://yomu.fyi/company/twilio.md)
- Author: Amanda Lange, Dylan Frankcom
- Published: Sep 14, 2026
- Type: Tutorial

The tutorial builds a Python FastAPI voice service that preserves caller context, preferences, and action history across separate inbound calls using Twilio Conversation Memory. It connects a Twilio phone number to a /voice webhook and /ws WebSocket through Conversation Relay, then uses an asynchronous OpenAI client with gpt-4o-mini to generate streamed responses. A memory service retrieves stored information by caller number, while each completed call is summarized and written back to the configured memory store. The example uses an auto-repair assistant whose system prompt incorporates prior context before the next call, and the test procedure checks that a second call recalls details from the first. The setup requires Python 3.11 or later, Twilio and OpenAI credentials, a memory-store and configuration ID, and ngrok for exposing local webhooks.


### [How to Connect Your Twilio Agent to External APIs in Node.js](https://yomu.fyi/post/how-to-connect-your-twilio-agent-to-external-apis-in-node-js.md)
- Company: [Twilio](https://yomu.fyi/company/twilio.md)
- Author: Dhruv Patel, Amanda Lange
- Published: Sep 14, 2026
- Type: Tutorial

Voice agents can converse, but they need external APIs to retrieve live information or take actions such as checking inventory or booking appointments. This tutorial builds a Node.js voice agent with Twilio Conversation Relay, Express, WebSockets, and the OpenAI Node.js SDK. A Conversation Relay WebSocket handler collects voice prompts, maintains conversation messages, streams model output back to the call, and defines a get\_cat\_fact tool that fetches JSON from the unauthenticated Cat Facts REST API. The service returns the fetched fact or an error message, while TwiML connects Twilio's voice-capable number to the WebSocket endpoint. Testing uses Node.js 18+, an OpenAI API key, ngrok, and a POST /voice webhook; the stated result is that callers can ask for a cat fact and hear the API response.


### [How to Orchestrate Multi-Call Conversations with an LLM and Twilio Conversation in Node.js Memory](https://yomu.fyi/post/how-to-orchestrate-multi-call-conversations-with-an-llm-and-twilio-con-nodejs.md)
- Company: [Twilio](https://yomu.fyi/company/twilio.md)
- Author: Amanda Lange, Dhruv Patel
- Published: Sep 14, 2026
- Type: Tutorial

Disconnected voice calls can force callers to repeat information, creating friction for AI phone agents. It builds a Node.js Express service that preserves caller context, preferences, and action history across separate inbound calls using Twilio Conversation Memory. The implementation uses Express and express-ws for HTTP and WebSocket handling, the OpenAI SDK with gpt-4o-mini for streamed responses and summaries, and axios for Twilio’s Conversation Memory REST API; a memory store ID and other credentials are configured through environment variables. The sample application connects Twilio’s Conversation Relay to an auto-repair assistant, retrieves prior context for recognized callers, and logs a summary when a call ends. Testing involves two calls from the same phone number to verify that the second call recalls details from the first.


### [Italy's email open tracking pixels rules are changing](https://yomu.fyi/post/italy-s-email-open-tracking-pixels-rules-are-changing.md)
- Company: [Twilio](https://yomu.fyi/company/twilio.md)
- Author: Denis O'Sullivan
- Published: Sep 14, 2026
- Type: Explainer

Italy’s data protection authority, the Garante per la protezione dei dati personali, has issued guidelines governing tracking pixels in email communications. The guidance says recipients must be told about tracking pixels regardless of the email’s purpose or sender, and that prior consent is required for many uses. Senders must comply by October 28, 2026; existing subscribers who signed up before April 29, 2026 need an opt-out opportunity, while subscribers from April 29 onward require prior, explicit consent for open tracking pixels. The guidance also includes complexities, including exemptions for some email types and different rules for opaque pixels, so senders are advised to review opt-in processes with legal or privacy teams. The change follows France’s earlier rules and may foreshadow similar action elsewhere in Europe.


### [Harbor: Stripe’s AI-assisted prototyping tool](https://yomu.fyi/post/harbor-stripe-s-ai-assisted-prototyping-tool.md)
- Company: [Stripe](https://yomu.fyi/company/stripe.md)
- Author: Cristian Rivera
- Published: Sep 14, 2026
- Type: Explainer

Harbor addresses a gap in Stripe’s AI-assisted workflows: designers needed dynamic information to assess stripe.com designs, but could not always justify a full engineering prototype. Built initially by one engineer in about 40 days, it lets designers describe an idea and use Stripe design systems and realistic mock data to generate shareable, browser-based prototypes without understanding code. Its agent writes a multi-file source tree, while Drydock compiles and renders React and TypeScript in the viewer’s browser; the agent can inspect pages, interact with elements, switch viewports, and iterate before publishing a single snapshot. Since May 2026, more than 3,000 Stripes have created 12,000 prototypes, and Harbor now integrates with other AI tools through MCP and Drydock’s standalone rendering contract, although live data and prototype-scoped state remain future goals.


### [How to Connect Your Twilio Agent to External APIs in Python](https://yomu.fyi/post/how-to-connect-your-twilio-agent-to-external-apis-in-python.md)
- Company: [Twilio](https://yomu.fyi/company/twilio.md)
- Author: Amanda Lange, Dylan Frankcom
- Published: Sep 12, 2026
- Type: Tutorial

Voice agents can converse, but need external APIs to retrieve live data or take actions such as checking inventory and booking appointments. This tutorial builds a Python and FastAPI voice agent with Twilio Conversation Relay, OpenAI tool calling, and an unauthenticated Cat Facts REST API. A FastAPI webhook returns TwiML that connects callers to a websocket, while Conversation Relay messages are converted into conversation history and streamed responses. When the model requests a cat fact, the application executes the tool, sends its result back to OpenAI, and streams the follow-up answer to the caller. The setup uses environment variables for the OpenAI key, httpx for asynchronous external requests, and ngrok to expose local webhooks; the completed test lets a caller ask for and receive a cat fact.


### [Engineering Real-Time Mobile Personalization Across iOS, Android, React Native, and Flutter](https://yomu.fyi/post/engineering-real-time-mobile-personalization-across-ios-android-react.md)
- Company: [Salesforce](https://yomu.fyi/company/salesforce.md)
- Author: Scott Nyberg
- Published: Sep 11, 2026
- Type: Explainer

Building real-time mobile personalization across iOS, Android, React Native, and Flutter required one architecture to handle different rendering models, lifecycles, identity flows, consent requirements, and release cycles without disrupting native app experiences. The team standardized schemas, event semantics, identity behavior, fallback handling, and SDK APIs, using bridge layers so React Native and Flutter could access native iOS and Android capabilities. Developers instrument apps once with constrained content zones and approved native components, while marketers configure templates, targeting, and content through Salesforce UI; metadata is published via a global CDN and rendered dynamically by the SDK. Data 360 performs server-side identity resolution and decisioning, limiting unnecessary raw profile storage on devices. An early customer implemented the system in two months and generated more than 150 million decisions within three weeks, while QR-code previews enabled testing against live profiles without app recompilation.


### [Health Plans: Your BI Tells You MLR Moved. Can Your AI Tell You Why?](https://yomu.fyi/post/health-plans-your-bi-tells-you-mlr-moved-can-your-ai-tell-you-why.md)
- Company: [Databricks](https://yomu.fyi/company/databricks.md)
- Author: Aaron Zavora, Jonathan Thompson
- Published: Sep 11, 2026
- Type: Explainer

Health plan finance teams can see when medical loss ratio (MLR) moves above budget, but the variance alone does not explain which business areas, costs, populations, or providers caused it. Conversational AI is presented as a way for finance leaders to ask follow-up questions across claims, membership, provider, contract, clinical, quality, risk, and financial data without repeatedly commissioning reports. Its usefulness depends on explicit payer context: MLR combines premiums with medical-care components such as pharmacy rebates, IBNR, recoveries, risk adjustment transfers, and provider settlements, while cohort definitions and metric logic vary. Abacus supplies normalized, connected healthcare data and payer-specific business context; Databricks provides governed enterprise data and AI capabilities for conversational analytical experiences. Together, they are described as creating a reusable foundation for tracing executive signals into operational questions across MLR and other payer domains, while leaving action judgment to the organization.


### [Unify your marketing data with Lakeflow Connect](https://yomu.fyi/post/unify-your-marketing-data-with-lakeflow-connect.md)
- Company: [Databricks](https://yomu.fyi/company/databricks.md)
- Author: Sonia Bendre, Giselle Goicochea
- Published: Sep 11, 2026
- Type: Announcement

Lakeflow Connect addresses the fragmentation of Customer 360 data across ad platforms, analytics tools, CRMs, and support systems. Its fully managed native connectors ingest sources through a UI or API, incrementally landing governed tables in Unity Catalog without infrastructure to provision, manage, or scale. Connectors for Google Ads, Meta Ads, TikTok Ads, Reddit Ads, and LinkedIn Ads provide pre-built or custom reports with spend, impressions, clicks, and conversions, while OAuth flows simplify setup for some platforms. The post also describes HubSpot, Salesforce, and other journey-wide connectors, plus the Ad-Genie accelerator, which transforms bronze data into silver and gold tables for AI/BI dashboards and Genie agents. Customer examples report reduced engineering overhead and faster ingestion.


### [From Signals to Context: Lessons from Lavaredo Ultra Trail](https://yomu.fyi/post/from-signals-to-context-lessons-from-lavaredo-ultra-trail.md)
- Company: [Uber](https://yomu.fyi/company/uber.md)
- Published: Sep 11, 2026
- Type: Explainer

The title “From Signals to Context: Lessons from Lavaredo Ultra Trail” presents a subject involving signals, context, and lessons associated with Lavaredo Ultra Trail. The available page content consists of Uber navigation and promotional text, including links for products, advertising, driving and delivery, Uber Eats, business, health, higher education, transit, engineering, community support, news, and Uber.com. It identifies Uber Blog as the surrounding publication area, but supplies no body text explaining what the signals are, how context is produced, or which lessons were drawn from the trail event. Consequently, no technical approach, implementation detail, measurement, result, or conclusion can be stated from the available content without adding information beyond the displayed page.


### [Evolving Pinterest’s Embedding Retrieval Platform](https://yomu.fyi/post/evolving-pinterest-s-embedding-retrieval-platform.md)
- Company: [Pinterest](https://yomu.fyi/company/pinterest.md)
- Author: Pinterest Engineering
- Published: Sep 11, 2026
- Type: Explainer

Pinterest describes how its Manas distributed search platform is evolving embedding retrieval for a corpus growing toward tens of billions of embeddings. Traditional HNSW indexes require all data in RAM, while single-vector two-tower retrieval limits context-dependent relevance, motivating changes across quantization, SSD-based serving, and multi-embedding retrieval. Manas uses Scalar Quantization and Product Quantization across HNSW and IVF, with SQ preserving over 90% recall while reducing index size, and Linear Scaling plus SIMD cuts per-query computing resources by 10–15%. For disk-based serving, SPANN with PQ is reported as 4.5x faster than plain SPANN and over 3x DiskANN QPS, while a five-billion-embedding production evaluation saved over 40% CPU time with under 5% recall drop. The platform is also adding ColBERT-style late interaction using Sum of MaxSim, with a pilot in progress.


### [From zero-shot forecast to purchase order with Amazon Bedrock AgentCore](https://yomu.fyi/post/from-zero-shot-forecast-to-purchase-order-with-amazon-bedrock-agentcor.md)
- Company: [AWS](https://yomu.fyi/company/aws.md)
- Author: Hyunsoo Kim, Ph.D.
- Published: Sep 11, 2026
- Type: Problem & solution

This post presents an inventory automation architecture that moves from zero-shot demand forecasts to validated purchase orders while applying safety stock, minimum order quantities, budgets, and promotional adjustments. It combines Amazon Chronos2, hosted on SageMaker Serverless Inference, with four Claude-powered agents built using the Strands Agents SDK and deployed on Amazon Bedrock AgentCore; deterministic Python tools perform computation and agents handle judgment. Product data and future covariates arrive through CSV files in S3, while business rules remain in JSON, enabling what-if forecasts and onboarding without per-SKU training. In internal testing across 50 SKUs over a four-week horizon, the system reached 12.3% median WAPE, reduced onboarding from 2–3 weeks to under five minutes, and lowered estimated monthly inference cost from about $1,091 to $15. A conditional retry loop surfaces budget violations for user decisions instead of silently truncating orders, with traces and evaluator scores supporting auditability.


### [Unlock a LUKS root over SSH on Fedora and Red Hat Enterprise Linux](https://yomu.fyi/post/unlock-a-luks-root-over-ssh-on-fedora-and-red-hat-enterprise-linux.md)
- Company: [Red Hat](https://yomu.fyi/company/red-hat.md)
- Author: Josephine Pfeiffer
- Published: Sep 11, 2026
- Type: Tutorial

Remote Fedora or Red Hat Enterprise Linux servers with encrypted LUKS roots can stop in the initramfs, waiting for a passphrase at an unattended console. The documented setup installs dracut-sshd from EPEL, adds dracut’s network-manager module, enables rd.neednet=1 and ip=dhcp in GRUB, then rebuilds the initramfs and boot configuration. NetworkManager brings up the initramfs network, while SSH exposes systemd-tty-ask-password-agent; the SSH session closes after unlocking because the root transition destroys the initramfs sshd. The rd.neednet=1 argument creates the neednet flag that gates nm-initrd.service, and the generated connection profile carries into the real root. Initramfs credentials sit unencrypted on /boot, so dedicated host keys, separate known\_hosts entries, and physical testing are recommended before relying on remote recovery.


[Newer posts](https://yomu.fyi/posts/page/45.md) · [Older posts](https://yomu.fyi/posts/page/47.md)