Loading…
Latest reads
Every post, summarised
Grounded summaries of engineering blogs, newest first. Each links to the original.
- 4,671 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
Introducing Genie Agent Mode
Databricks introduces Agent mode in Genie spaces, an agentic process that plans, explores, and reasons over data to answer business questions. For each question, Genie can confirm an issue, test potential contributors through multiple queries, reflect on results, and decide what to investigate next, using Unity Catalog metadata and author-defined semantics from the Genie space. It produces a findings report with quantified results, identified contributors, visualizations, and references to the underlying SQL, and may include actionable recommendations. The agent scales its reasoning to task complexity, using validation for straightforward questions and more planning for multi-step investigations; in AI/BI Dashboards, Genie uses Agent mode by default. Workspace admins enable the feature in Workspace Previews, after which users turn on the Agent toggle in Genie spaces; API support and unstructured document analysis are planned additions.
Governing coding agent sprawl with Unity AI Gateway
Databricks introduces coding agent support in Unity AI Gateway to address security, cost, and visibility challenges created by organizations using multiple coding tools. The gateway provides a unified governance hub for Codex, Cursor, and Gemini CLI, combining access controls, usage statistics, operational observability, cost management, guardrails, and inference capacity. Agent access can be governed centrally, with audit logs in Unity Catalog, MCP servers managed in Databricks, MLflow tracing, shared cost limits, and coding metrics and traces ingested into Unity Catalog-managed Delta tables through OpenTelemetry. The announcement says developers can authenticate with Databricks credentials across connected services, use centralized budgets and model capacity, and let organizations analyze adoption, developer velocity, and rate-limit pressure in the lakehouse; support for Cursor, Gemini CLI, and Codex CLI is available to all Databricks customers.
otel.fyi - OTel Collector docs made simple
OpenTelemetry Collector documentation is often fragmented across separate contrib repositories as individual README files, making component configuration difficult and time-consuming to navigate. To address this friction, the team created otel.fyi as a dedicated platform to aggregate, simplify, and structure documentation from across the ecosystem. The implementation parses contrib repositories to pull configuration examples and normalize metadata across receivers, processors, exporters, and extensions into a unified format. By surfacing structured component information upfront, the site allows practitioners to assemble working configurations rapidly without digging through multiple GitHub repositories. Originally developed for internal operational workflows, the platform has been released publicly to assist the broader community, with future plans to introduce an AI assistant.
“A generational leap”: How Trio unified payment analytics and cut storage by 88% with ClickHouse Cloud
Brazilian payment processor Trio faced severe data challenges after scaling and acquiring PayBrokers, processing over 243 million payments in early 2025 across disparate legacy databases. Because financial reconciliation and compliance demand absolute precision, the engineering team unified their analytics layer by migrating to ClickHouse Cloud. To handle out-of-order and duplicate events from their Redpanda streaming pipelines, Trio implemented sliding-window refreshable materialized views alongside ReplacingMergeTree engines rather than recomputing entire historical datasets. A custom ETL service manages schema drift before inserting clean records into ClickHouse, while batch ingestion is coordinated through Airflow. Consequently, Trio reduced storage consumption by approximately 88% through columnar compression, ingested a 5-billion-row migration without issue, and maintained real-time query performance across more than one billion daily ledger data points.
What is pgvector?
pgvector is an open-source PostgreSQL extension that adds vector storage, indexing, and similarity search for embeddings alongside relational application data. It supports semantic search, recommendations, retrieval-augmented generation (RAG), image similarity, anomaly detection, and deduplication without requiring a separate vector database for many workloads. The extension uses the vector data type and distance metrics including L2, cosine similarity, and inner product, with HNSW and IVFFlat indexes offering different speed and memory trade-offs. Vector queries can also combine relational filters with Postgres full-text search for hybrid retrieval. pgvector is positioned as a practical choice for existing Postgres deployments, while pgvectorscale can extend its usefulness as data and query demands grow; dedicated vector databases become more relevant at larger scales, and Databricks AI Search serves complementary lakehouse workloads.
Banks don't have an AI problem — they have a data platform problem
The post argues that banks’ difficulty scaling AI stems less from model capability than from fragmented data, weak governance, and limited real-time access. Examples from CBA Live 2026 include model drift in credit scoring, fragmented customer signals, and a collections model that predicted, with 85% accuracy, how many days newly delinquent accounts would take to cure using governed data. It also presents Erica’s 3.2 billion interactions since 2018 as evidence that production AI requires continuous tuning, monitoring, and edge-case management, while generic models decay in frontline settings. The proposed Databricks architecture combines Lakehouse, Unity Catalog, Lakeflow, Lakebase, MLflow, Model Monitoring, Online Feature Store, Genie, and Agent Bricks to support governed analytics, low-latency decisions, auditable models, and constrained agent actions. Its conclusion is that a shared data platform should precede additional AI use cases because it improves deployment speed, trust, explainability, and regulatory defensibility.
React Email 6.0
React Email 6.0 announces an open-source visual editor as its centerpiece, available as a standalone package that outputs email-ready HTML for major inbox providers. The editor is built in two layers: a core that works without configuration and a composable API using EmailNode for custom extensions, with optional themes for styling. The release ships new authentication and e-commerce templates from Character Studio, available as React Email templates or Figma files. All components are now unified into a single package except the editor, simplifying imports and reducing package management. The project has reached 2 million weekly downloads, a 108% increase since the last major release five months ago.
Open Platform, Unified Pipelines: Why dbt on Databricks is Accelerating
dbt provides structure for transforming raw data into curated datasets, but the post argues that its effectiveness depends on the underlying platform. Databricks is presented as an open lakehouse combining storage, compute, governance, orchestration, and monitoring to reduce duplicated data, fragmented permissions, and operational overhead. Lakeflow Jobs runs dbt alongside ingestion, downstream actions, retries, and dependencies, while Unity Catalog supplies persistent permissions, documentation, lineage, and fine-grained controls. The integration also supports Delta Lake and Apache Iceberg, query tags for cost tracking, Photon, Predictive Optimization, Liquid Clustering, and dbt materialized views. The post concludes these capabilities provide openness, integrated operations, governance, and price-performance for dbt workflows, citing one customer’s consolidation of notebooks and 7+ source systems and more than 2,900 customers already running dbt on Databricks.
Building a Unified Pipeline for AI Token Spend
AI token spend is volatile, consumption-based, and distributed across teams, making usage and cost difficult to monitor. Ramp's pipeline ingests LiteLLM and OpenRouter events through authenticated, tenant-scoped webhooks, Kafka, and ClickHouse, then aggregates them for REST dashboards and AI-generated forecasts. LiteLLM uses callbacks with token counts and metadata, while OpenRouter sends OTLP traces; ReplacingMergeTree deduplicates replayed events using business_id, source, and event_id. Customers can connect gateways with an API key, configure LiteLLM callbacks or OpenRouter Broadcast, and pass attribution metadata. The resulting visibility supports breakdowns by provider, model, department, user, project, and use case, and exposed phantom Gemini reasoning tokens after a LiteLLM upgrade, geo pricing premiums, runaway loops, and prompt bloat.
How GitHub uses eBPF to improve deployment safety
Deployment scripts can introduce dangerous circular dependencies when they rely on services or assets from platforms that are currently experiencing outages. Blocking network access at the host level is impractical because stateful nodes continue serving live traffic during rolling deployments. To solve this, GitHub isolates deploy scripts into dedicated Linux cGroups and attaches custom eBPF programs via the cilium/ebpf Go library. The system uses socket-address hooks to redirect DNS queries to a userspace proxy that checks a domain blocklist, while egress packet hooks map DNS transaction IDs to process IDs. This approach successfully prevents deploy-time circular dependencies, provides full command-line audit logs for blocked requests, and speeds up incident recovery.
April 2026 newsletter
ClickHouse 26.3 LTS introduces experimental materialized common table expressions, WebAssembly user-defined functions, and enables asynchronous inserts by default. Alongside the release, Alasdair Brown launched clickhousectl to manage local installations and cloud deployments with built-in agentic development features and authentication safeguards. User case studies show Goldsky accelerating historical blockchain backfills by twelve times through direct ClickHouse reads using Apache Arrow, replacing their prior Kafka and Avro architecture. Additionally, the engineering team redesigned the full-text search index with front-coding compression, a sparse lookup index, and adaptive posting lists to handle expensive random reads on object storage. ClickHouse also previewed community architecture patterns for lightweight SIEM setups using Clickdetect and Wazuh log collection.
Why your agents can't read enterprise documents — and how to fix it
Enterprise agents often reason well over clean text but struggle with scanned PDFs, inconsistent layouts, nested tables, images, and handwriting, making document processing an accuracy ceiling for business workflows. The post announces Document Intelligence, a pipeline of chainable AI Functions: ai_parse_document converts scans into layout-enriched structured text, while ai_classify and ai_extract support routing and entity extraction. On OfficeQA treasury-bond documents, preprocessing with ai_parse_document produced a 16% average performance gain across tested agent frameworks without changing their reasoning harness. Across invoice, contract, medical-note, and financial-filing extraction tasks, it reportedly delivered the highest accuracy at 5–7x lower cost than comparable pipelines, with serverless batch infrastructure scaling the same SQL call from 100 to 100,000. Databricks integrates ingestion, orchestration, governance, and agent building through Lakeflow Connect, Lakeflow Jobs or Spark Declarative Pipelines, Unity Catalog, and Agent Bricks, aiming to replace fragmented services with reusable workflows.
Building with Databricks Document Intelligence and Lakeflow
The post presents a production-oriented approach to Intelligent Document Processing (IDP) that brings PDFs, images, and office documents into Databricks’ governed data lifecycle. Lakeflow Connect ingests files from sources such as SharePoint and Google Drive into Unity Catalog Volumes and tables, applying access control, lineage, and auditing while supporting incremental reads and writes. Databricks’ Document Intelligence then uses ai_parse_document to preserve structure in VARIANT output, with ai_extract, ai_classify, and ai_prep_search available for enrichment, routing, and retrieval preparation. Lakeflow Jobs productionizes the workflow with orchestration, control flow, retries, triggers, serverless scaling, and observability across ingestion, parsing, and serving tasks. The post also recommends joining extracted data to master data, using service principals, and monitoring drift and extraction accuracy, presenting the combined platform as a way to create trusted, queryable datasets from unstructured content.
Databricks on Google Cloud: Innovate faster. Smarter. Together.
Databricks and Google Cloud are presenting their partnership at Google Cloud Next ’26 as a route from governed enterprise data to production AI, including autonomous agents and modernized workloads. In 2025, Databricks became a first-party provider of Gemini models, giving users direct API access to build, customize, and deploy agentic AI on governed data across all three Databricks clouds without moving data or managing infrastructure. The post reports more than 55% quarter-over-quarter growth in Gemini adoption, 85% year-over-year growth in Google Cloud consumption, and 4,000-plus Marketplace sign-ups, alongside support for Google Axion processors. Customer examples include Databricks SQL Serverless savings for Digital Turbine, faster Marketplace procurement for Dun & Bradstreet, regional compliance in Saudi Arabia and Brazil, and 30% faster analytics delivery through Lakeflow-powered pipelines. At the event, Databricks is offering demos, talks, expert sessions, and a free $400 trial through Google Cloud Marketplace.
What is Generative AI in Marketing?
Generative AI in marketing creates content, insights and recommendations that can personalize experiences, optimize campaigns and improve performance. Unlike traditional analytics, which mainly reports on past results, it produces net-new outputs such as ad copy, audience segments, product recommendations, visual assets and strategic summaries. A typical workflow prepares campaign, customer and brand data, grounds or fine-tunes models, generates outputs, applies targeting and optimization, and uses human review and refinement. It distinguishes pretrained tools such as ChatGPT, Claude and Perplexity from customized models and broader AI transformation, noting trade-offs between speed, investment, relevance and strategic alignment. Successful adoption depends on high-quality, consented data, governance, privacy controls, human oversight, monitoring and cross-functional ownership; the conclusion presents AI as a force multiplier rather than a replacement for human creativity and judgment.
Custom Tracking Domains
Resend has launched custom tracking domains, making the feature free for all users. Email tracking typically relies on shared domains, so another sender’s reputation can affect deliverability. Users configure the feature at the domain level by adding a required DNS record, after which Resend routes tracking through the custom domain; both click and open tracking are disabled by default. Click tracking records link clicks through redirects, while open tracking records image downloads from a 1x1 transparent pixel. Custom domains align tracking links with the sending domain, avoid unfamiliar recipient-facing domains, and can improve deliverability; API configuration accepts open and click tracking flags plus a tracking subdomain, while webhooks emit real-time open and click events.
Ecom-RLVE: Adaptive Verifiable Environments for E-Commerce Conversational Agents
Ecom-RLVE extends the RLVE framework from single-turn reasoning puzzles to multi-turn, tool-augmented e-commerce conversations. EcomRLVE-GYM supplies eight environments spanning product discovery, substitution, cart building, returns, order tracking, policy QA, bundle planning, and multi-intent journeys, with procedural problem generation, a 12-axis difficulty curriculum, and algorithmically verifiable rewards. Each episode requires tool use and world-state changes; rewards combine task completion, efficiency, and a penalty for recommending product IDs not retrieved, while malformed outputs and illegal calls fail immediately. Adaptive scheduling advances each environment independently when its success rate is reliable. An initial study trained Qwen 3 8B with DAPO for 300 steps on cart building and reported progressively increasing difficulty reached, consistent with a steady learning signal.
The PR you would have opened yourself
mlx-lm contributors can use a Skill and a separate test harness to port language models from transformers, with the goal of making new architectures available shortly after their upstream implementations land. The work addresses rising agent-generated pull-request volume and the difficulty of preserving implicit codebase conventions, user contracts, and performance-sensitive behavior that agents may not understand. Given a conversion prompt, the Skill sets up environments, finds model variants and checkpoints, reads transformers modeling code, writes an MLX implementation, and iterates through tests including dtype checks, generation examples, numerical comparisons, and per-layer comparisons. The resulting pull requests disclose agent assistance and include detailed verification data, while contributors remain responsible for reviewing and accepting the results. Current limitations include shared-utility refactors, vision-language models, quantized uploads, and thinking-specific tests.
Training and Finetuning Multimodal Embedding & Reranker Models with Sentence Transformers
Sentence Transformers provides a Python library for training and using embedding and reranker models across text, images, audio, and video in retrieval and semantic search applications. The post shows how to finetune multimodal models on task-specific data, using Visual Document Retrieval as the example: matching text queries with document page images while preserving charts, tables, and layout. It describes the SentenceTransformerTrainer pipeline, model and processor configuration, dataset handling, and alternative Router-based encoders, while also covering multimodal reranker training with BinaryCrossEntropyLoss and direction-specific datasets. On its evaluation data, finetuning Qwen/Qwen3-VL-Embedding-2B raised NDCG@10 from 0.888 to 0.947 and outperformed all tested VDR models, including models up to 4x its size.
How we built Stripe Credits: A programmable, auditable way to pay your Stripe fees
Stripe Fee Credits replaces ad-hoc discounts and manual cash adjustments with a programmable, auditable virtual balance dedicated to eligible Stripe fees. The architecture has three layers: a standalone gRPC credits engine for metadata, policies, lifecycle, and audit history; logical financial accounts built on Money Movement and Storage; and fee-settlement orchestration that applies eligible credits before the primary balance. Logical accounts carry ledger balances without representing real cash, while platform checks restrict Originated Money Transfers to logical-to-logical movements and exclude these accounts from cash sweeps. Prepaid and Stripe-funded credits share the user experience but use separate funds flows: prepayments create deferred revenue and collect tax at drawdown, whereas Stripe-funded credits are booked as Stripe costs with taxes written off. The resulting infrastructure supports incentives, prepaid annual fees, and other programs across 26+ countries, with Dashboard balances and real-time auditability, while avoiding bespoke engineering for each commercial model.
Top topics
- Architecture834
- Performance816
- LLMs660
- Machine Learning622
- Open Source563
- AI Agents553
- AI470
- Developer Experience448
Weekly digest
The best engineering reads, once a week.