Loading…
Latest reads
Every post, summarised
Grounded summaries of engineering blogs, newest first. Each links to the original.
- 4,668 posts
- 20 companies
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
Latest posts
Bring Databricks into Kiro IDE with the AI Dev Kit Power
AI-assisted development can produce unreliable SQL and models when an assistant guesses schema details or exceeds the user's data access. Kiro can connect to Databricks through Model Context Protocol (MCP) in two ways: four Databricks-managed remote servers for Genie, SQL, Unity Catalog Functions, and Vector Search, or the Databricks AI Dev Kit Power, which installs a local Python MCP server and broader skills. The AI Dev Kit now supports Kiro through its unified installer, while the Power provides one-click onboarding with authentication detection and skill loading. Both paths use Unity Catalog permissions, including row-, column-, and tag-based grants, so the assistant sees the user's effective access; Path A uses token-based configuration, while Path B supports OAuth U2M, OAuth M2M, profiles, or PATs. Examples show schema-grounded SQL, dbt joins using real columns, query comparisons, lineage checks, and generation of Databricks jobs or Asset Bundles.
amazon.science
Ground truth is a process, not a dataset
Evaluating search-augmented AI research reports requires verifying complex, multi-source claims that depend heavily on surrounding context. Unassisted PhD-level domain specialists tasked with establishing static ground-truth labels achieved only 60.8% accuracy on a hidden test set during controlled testing. To resolve benchmark errors on cognitively demanding tasks, Amazon AGI introduced the audit-then-score evaluation protocol alongside DeepFact-Bench and the DeepFact-Eval verification system. Under this protocol, model disagreements serve as challenges requiring concrete counterevidence and written rationales, enabling human auditors to evaluate competing claims and update benchmark ground truth prior to scoring. Operating under this iterative workflow, expert test accuracy rose to 90.9% across four rounds, and DeepFact-Eval reached 83.4% accuracy with GPT-4.1.
Coding Is No Longer the Constraint: Scaling Developer Experience to Teams and Agents at Spotify
Spotify experienced rapid codebase growth alongside widespread adoption of AI coding tools, resulting in a 76% increase in pull request frequency. To manage fleet-wide codebase maintenance and complex refactoring, the company combined its Fleetshift automation platform with Honk, a background coding agent powered by Claude. Honk runs inside Kubernetes pods with access to CI build environments and integrates with Backstage via Model Context Protocol tools to inspect component metadata. Standardized service architectures and active lint guardrails enable the agent to maintain consistency and auto-correct invalid design patterns during execution. By automating routine migrations, Spotify reduced fleet-wide Java backend updates from multi-week cross-team efforts down to three days handled by a single engineer.
Introducing the Services Track and Partner Hub of the Claude Partner Network
Anthropic announces the Services Track and Claude Partner Hub to make its partner ecosystem easier for enterprise customers to evaluate firms that can put Claude into production. The Services Track has Select, Preferred, and Global Premier tiers, measured by certified practitioners, production deployments, public customer stories, and for Global Premier, a joint business plan with named executive sponsors. Requirements include at least 10, 100, or 1,000 active certified individuals, alongside tier-specific deployment and endorsement thresholds, with a single bar for firms of all sizes. The Hub provides daily refreshed standing dashboards, a public directory showing each firm's tier and evidence, and an MCP connector that lets partners query partnership information through Claude. Promotions occur twice yearly, with an additional 2026 review, while downgrades are limited to the annual review after notice and an opportunity to close gaps.
Direct Preference Optimization Beyond Chatbots
DharmaOCR applies Direct Preference Optimization (DPO) beyond chat alignment to reduce text degeneration in structured OCR, where outputs are either acceptable transcriptions or failure loops. In a benchmark of Brazilian Portuguese structured-document OCR, vanilla degeneration rates among open-source model families ranged from below 1% to above 33%, while supervised fine-tuning (SFT) did not consistently reach production-acceptable levels. After SFT, the pipeline generated multiple candidates for 23,726 documents, used an automated LLM judge to score them, and retained the model’s degenerate outputs as rejected examples paired with chosen outputs. DPO reduced degeneration in every tested family, by 59.4% on average and up to 87.6%, without reducing extraction quality according to the benchmark; applicability is presented as dependent on failures being categorically identifiable, reliably scoreable, and sufficiently numerous.
Introducing organizations for Cursor Enterprise
Cursor is introducing organizations, a top-level structure that lets Enterprise customers manage multiple teams, budgets, security policies, governance, and feature controls from one dashboard. Organizations contain teams for departments or subsidiaries and groups for user cohorts; separate configurations can govern model access, spending, agent permissions, and network access, while the most permissive setting applies when users belong to multiple teams or groups. The organization dashboard aggregates spend and token usage across teams and supports filtering by team, user, service account, or cloud agent, while identity-provider and SCIM setup is centralized and membership can be managed through the dashboard, API, or CSV. Generally available to all Enterprise customers, the structure supports sandbox teams for testing features, scoped administration and chargebacks, and differentiated controls across functions; Cursor says policy controls, onboarding, and SCIM-driven assignment are still being expanded.
Stack Benchmarking
Ramp’s Stack is an AI-native accounting suite designed to automate repetitive book-closing work, including reconciliations, variance analysis, data entry, and schedules and accruals. To avoid overfitting to individual design partners, Ramp built a benchmark of synthetic business worlds, realistic accounting tasks, and accountant-written grading criteria, including standard and roll-forward worlds for testing memory transfer. The benchmark contains 237 tasks and 3,469 grading criteria in the analyzed slice, and supports repeated runs to compare models, prompts, tools, skills, harness changes, and memory behavior. Optimization included ablating skills, shrinking a spreadsheet skill from 14,000 to 5,000 characters, and tuning the system against end-to-end task performance rather than narrow evaluations. The resulting Stack system achieved the highest agent performance, with 4% higher accuracy and 3% better Pass@1, while remaining on the latency frontier with GPT 5.4; schedules and accruals were harder for raw models than variance analysis.
Scaling Enterprise Conversational Intelligence: Cross-industry Technology and Functional Solutions Powered by Databricks Genie
Databricks Genie is presented as a cross-industry technology layer for enterprise challenges including financial planning, legal compliance, and IT operations. As a “Research Agent,” it can generate multi-step research plans to explain business anomalies and support answers with verifiable proof from the lakehouse. The post showcases partner solutions across technology, sales, marketing, HR, finance and procurement, supply chain, customer service, and IT operations, with examples spanning governed analytics, multi-agent orchestration, data observability, causal analysis, and incident management. These implementations aim to replace fragmented or static workflows with real-time, contextualized intelligence and production-grade agentic workflows, supporting anomaly investigation, root-cause analysis, ticket classification, and conversational troubleshooting. The stated goal is faster, more confident decision making across departments through governed self-service access to insights.
Logging, Metrics, and Distributed Tracing in .NET with OpenTelemetry and ClickStack
The post presents a walkthrough of instrumenting two ASP.NET services—an Order API and a Payment Service—with OpenTelemetry and sending traces, logs, and metrics to ClickStack. The demo uses SQLite for persistence, with Entity Framework Core instrumentation capturing database operations without manually creating spans. The configuration adds ASP.NET, HTTP client, and EF Core instrumentation, custom sources and meters, OpenTelemetry logging, and OTLP/gRPC exporters directed by OTEL_EXPORTER_OTLP_ENDPOINT. In ClickStack, the signals are correlated so a request can be followed across service boundaries, application work, and database calls, while the UI and ClickHouse SQL provide exploration options. The conclusion is that a relatively small ASP.NET setup replaces isolated timeout logs with an end-to-end view of timing, errors, logs, and database activity.
Organizations for Cursor Enterprise
Cursor has made Organizations generally available to Enterprise customers, enabling them to manage multiple teams from one place. An organization is the top-level container for company identity, administration, membership, and rollups of spend and token usage across teams. Teams remain the operating units for departments, regions, or subsidiaries, but now sit under an organization and can have distinct security, governance, spend, and feature settings. Groups provide lightweight user cohorts across or within teams, with separate model access, spend limits, and agent permissions. A user may belong to multiple teams or groups, and the most permissive setting wins; existing customers retain their current team as the default home for login, routing, and new team creation.
3,000,000 users
Resend reports reaching 3,000,000 users six months after reaching 1 million, while its team grew from 25 to 43 members and weekly npm downloads increased from 1.4M to 6.8M. The company says it has been profitable since last year without reducing cloud or Datadog spending or slowing hiring, treating profitability as flexibility for product decisions rather than the goal. Its support model gives AI agents the product’s tools for routine questions while keeping engineer support available when answers are not found. For developers and agents, Resend has made every page render as Markdown, expanded MCP coverage from about 8% to about 90%, shipped a 53-command CLI, and planned remote MCP and OAuth, while positioning Automations alongside sending, receiving, and broadcasting in one email platform.
Bringing Gemma 4 12B to your Laptop: Unlocking Local, Agentic Workflows with Google AI Edge
Google DeepMind’s Gemma 4 12B is presented as an open model for running agentic, multimodal capabilities locally on everyday laptops through the Google AI Edge stack. The release covers Google AI Edge Gallery for local data analysis and coding, including Python generation, execution, chart rendering, and a reported 3D rendering workflow that can generate dependencies, code, and self-corrections in one turn. Google AI Edge Eloquent adds fully on-device dictation, transcription, and Voice Edit commands for rewriting or translating selected text, while LiteRT-LM’s new serve command exposes the model through a local, OpenAI-compatible endpoint for tools and frameworks. The post says Gemma 4 12B delivers a 60%+ overall quality improvement over prior models for Eloquent’s use case and keeps data on-device, with performance and memory benchmarks provided in the model card.
Gemma 4 12B: The Developer Guide
Gemma 4 12B is a dense multimodal model with a unified, encoder-free architecture, native audio input, and a size intended for local execution on dedicated GPU laptops with 16GB VRAM or unified memory. Instead of separate vision and audio encoders, it uses a 35M-parameter vision embedder that projects raw 48x48 patches and a linear projection for raw 16 kHz audio sliced into 40ms frames; both feed a decoder-only transformer with shared weights. That design reduces multimodal latency and fragmented memory footprints, while allowing downstream adapters such as LoRA or full tuning to update the multimodal token loop in one pass. Capabilities include automatic speech recognition, agentic reasoning, diarization, video understanding, and coding, with examples covering a locally built image-processing app and five minutes of video processed at 1 FPS with audio; LiteRT-LM adds offline macOS apps, sandboxed Python execution, and an OpenAI-compatible local API server with stateless prefix caching.
Adding MCP Tools to Reachy Mini
Reachy Mini’s conversation app can now call tools hosted in public Hugging Face Spaces through MCP, allowing capabilities such as weather checks and web search to be added without editing the app or downloading tool code locally. The app retains built-in robot tools and custom local tools, while profiles use tools.txt to control which tools the model may call. The tool-spaces command validates a Space, probes its MCP endpoint, discovers tool IDs, and can append them to a selected profile; remote names are namespaced to avoid collisions. The feature supports public MCP-compatible Gradio Spaces, multiple Spaces, and backend-agnostic registration, but not private or authenticated Spaces, non-Gradio Spaces, arbitrary MCP URLs, or guaranteed parallel orchestration.
How ClickHouse became fast at joins
Over two years, ClickHouse focused on making joins faster for analytical SQL workloads, moving them from a performance gap to a first-class strength. The first year established the foundation with parallel hash joins, local automatic join reordering, filter pushdown, and execution improvements, producing a 4.4× speedup on the TPC-H SF100 join-heavy workload by version 25.4, compared with 22.4. The second year made optimizations effective by default, adding correlated-subquery support, lazy column replication, runtime filters, and statistics-based join reordering. Correlated queries can now run directly, while compact representations avoid copying repeated values and runtime filtering skips rows before hash-table lookups. Across the same benchmark and hardware, these changes delivered another 6× speedup, for a total 26× improvement; one six-table query fell from 3903.7 seconds to 2.7 while peak memory dropped from 99.1 GiB to 3.9 GiB.
Uber Advertising Launches New APIs to Unified Campaigns & Workflows
Uber has launched Uber Ads APIs to reduce the need to manage advertising and performance across separate tools and systems. The APIs let POS systems, aggregators, agencies, and other partners create and manage campaigns programmatically while giving brands direct or integrated access to performance and sales data. They support onboarding, activation, and ongoing campaign management across a merchant base, including campaigns for restaurants and Sponsored Listings. Partners can embed Uber Ads in existing platforms so merchants can launch campaigns and increase visibility without navigating separate workflows. Documentation, representative contact, and an integration-planning process provide the stated starting points, with the broader aim of combining advertising and operations in a unified retail media system.
Beyond parsing X12: Closing the gap for revenue cycle workflows in healthcare
Healthcare X12 pipelines can normalize remittance and enrollment data without giving medical billers a place to act on denials, leaving short-pays in SQL queries and spreadsheets while filing deadlines approach. Genpact and Databricks address that workflow gap with a two-layer operational workbench inside Databricks: production extensions to the open-source x12-edi-parser and a secure web application over Databricks SQL. The engine adds contextual 834 enrollment tracking and decoded 835 CAS adjustments, while the interface provides claim and service-line drill-downs, denial queues, dashboards, quality gates, and appeal or correction drafting. Live Unity Catalog gold-view queries avoid ETL shadow copies and synchronized caches; PHI remains inside the secure perimeter, with row-level security and audit logging for reveals and workflow changes. The proposed next step is Claude-based appeal drafting through Databricks Foundation Model APIs, and a two-week configuration-based trial uses an organization's schema and twenty existing denials.
Agentic BI: A Practical Guide for BI Teams and Business Users
Agentic BI uses autonomous AI agents to automate work between raw business data and actionable insight, including data preparation, query execution, chart and narrative generation, and report distribution. Traditional BI depends on analysts to gather data, write queries, maintain dashboards, and assemble reports, while agentic systems let business users ask natural-language questions and receive governed answers. The guide identifies a governed semantic layer as foundational, because shared metric definitions and deterministic execution help keep outputs consistent, auditable, and trustworthy, with human approval checkpoints for higher-risk handoffs. It recommends inventorying data structure, schema drift risk, and integration costs, then piloting a narrowly defined workflow and measuring time to insight, analyst hours reclaimed, satisfaction, and accuracy before expansion.
When history fails you, borrow from geography
When historical travel data became uninformative during the asynchronous post-COVID recovery, Airbnb redesigned its demand forecasting to borrow signals across geographies rather than waiting for local data. Using a hierarchical Bayesian framework, the team propagated updated posterior estimates from early-recovering corridors as informative priors for structurally similar, later-recovering corridors. This allowed accurate corridor-level demand forecasting in near real time during periods of high disruption and uneven market reopening.
Data Science vs Data Analytics: Compare Careers, Skills, and Degrees
Data analytics and data science are compared as data-focused career paths, with analytics interpreting existing, usually structured data for trends and business decisions, while science builds models and automated systems to predict future outcomes. The guide contrasts typical questions, outputs, tools, education paths, and roles: analysts use SQL, Excel, Tableau, or Power BI for dashboards and reports, whereas data scientists use Python, R, Apache Spark, and MLflow for predictive models and algorithms. It describes analytics types from descriptive through prescriptive, and a data science workflow spanning collection, feature engineering, training, validation, and deployment, including unstructured data such as text, images, and sensor streams. It also explains collaboration, including analysts defining problems and baselines before scientists build models, and offers portfolio suggestions and questions for choosing a path.
Top topics
- Architecture834
- Performance815
- LLMs659
- Machine Learning621
- Open Source563
- AI Agents553
- AI469
- Developer Experience448
Weekly digest
The best engineering reads, once a week.