Alex Ott

Hallucinations are a systems bug, not a model bug. We built a system-level runtime layer Ghostlayer ,that wraps an existing open 20B language model. The base model was not changed. No fine-tuning. No retraining. No quantization. Same model. Same weights. Same prompts. The only difference: an inference-time system that listens to uncertainty and routes trust dynamically. All improvements come entirely from the system layer (GhostLayer + Ghost Routing). This is not a post-processing filter. This is not a pre-processing trick. This is an inference-time control layer. The core claim: hallucinations are primarily a runtime systems problem — not a model capability problem. We call this layer Ghostlayer. A reliability and routing layer for large language models. 👉 jde.ghostrouter.fi (technical methodology + benchmark

What is the replication tax and how does Diskless slash cross-AZ egress costs? This technical analysis by our Director of Data Streaming Engineering - Josep Prat, shows how you can remove the replication tax associated with Apache Kafka and build a more efficient streaming platform. https://lnkd.in/e5w6vQfx #ApacheKafka #Kafka #CloudCost #diskless 

8

Still configuring cache with complex #VCL logic? We’ve been there. And we built something better... #Trident uses a clean, declarative TOML config - not a tangle of abstract syntax nobody wants to debug on a Friday afternoon. Why it matters for developers: ✅ Zero “magic syntax” to learn ✅ Easy to read and onboard across dev / ops / devops ✅ Version-controllable just like any other config in your repo ✅ Built for modern CI/CD workflows Simple. Transparent. Maintainable. No more guessing what the cache will do. You define it. Clearly. #Trident is fully compatible with: Magento Open Source, Shopware, WooCommerce, PrestaShop, Sylius Launching Feb 14 Follow us for daily previews — #TridentIsComing

8

⚡Hey .NET Friends – PostgreSQL Just Joined the FunkyORM Party... After a few decades of wrestling .NET data-access layers (and watching frameworks come and go), we at Funcular Labs still believe the best tool is the one you barely notice... the one that just works. That's why we're genuinely excited to announce full PostgreSQL support in Funcular.Data.Orm—better known around here as FunkyORM—now live in version 3.1.0. You already know the drill if you've used our ORM with SQL Server: zero-configuration POCO mapping, lambda-powered LINQ queries that generate clean parameterized SQL, remote properties that flatten your object graphs without the N+1 tax, and performance that often leaves heavier ORMs in the dust. No DbContext boilerplate. No XML mapping files. Just your plain classes and the queries you’d write anyway. Now the same delightful API works seamlessly with Postgres. Our new PostgreSqlOrmDataProvider (built on Npgsql) handles all the dialect differences transparently—double-quoted identifiers for reserved words, LIMIT/OFFSET paging, RETURNING clauses on inserts, native BOOLEAN columns... everything you’d expect. You can literally swap providers and keep your entity classes untouched (yes, even the ones with [RemoteProperty] and [RemoteKey] magic). Whether you’re starting fresh or migrating an existing FunkyORM project, the experience is identical. And because we’re still the same lightweight micro-ORM at heart, you get the speed you love without giving up type safety or developer happiness. Want to give it a spin? Head over to the NuGet package and pull in the latest bits: https://lnkd.in/gMfmJWhc We’d love for you to try it out on your next (or current) Postgres project. Drop the provider in, point it at your database, run a few queries... then swing by the comments and tell us how it feels. Did it save you time? Make your code cleaner? Surprise you in a good way? We read every note. Here’s to simpler, faster data access—no matter which database you call home. #dotnet #csharp #postgresql #orm #microorm #FuncularLabs

2

#Redis is known as a fast #caching tool, but can fall short when #data persistence and durability are required. Redis can persist data, but it's an add-on, not a foundational feature. That means many trade-offs, including: - Higher latency - Large storage volumes - Limited scaling - Confusing configurations Users looking for hyperfast speeds and comprehensive data persistence don't have to settle for one or the other. #Aerospike is architected from the ground up for high throughput, ultra-low latency, and native persistence. Learn about Redis’ technical realities and how Aerospike is a better persistence #cache alternative in our latest blog: https://lnkd.in/gGMgFrmg #RedisReplaced #DatabaseEngineering #BackendDevelopment #Developers #TechStack #DataDurability

16

Redis announced two new products: LangCache for caching LLM responses and vector sets, a new data type in Redis for storing and querying embeddings. By Frederic Lardinois

1

In Structurizr news, an enhanced version of the "architectural perspectives" feature is shipping in the next release (likely at the weekend, but available now if you use the structurizr/structurizr:preview Docker image). This feature provides an implementation of the perspectives concept from "Software Systems Architecture" by Eoin Woods and Nick Rozanski, allowing you to add additional metadata to your software architecture model and view it via an "architectural perspective". The screenshot below shows a collection of software systems, each with some metadata about the level of technical debt, toggled on the UI by enabling the "Technical Debt perspective". There are two new features of note: 1. Perspectives can now have their own styling, so you can change the colour of boxes and arrows according to the perspective and it's value. 2. Perspectives can also be "dynamic", where the value of the element/relationship perspective is retrieved via a HTTP GET request in the browser. This makes it easy to perform simple health checks from your architecture diagram, for example. I hope you find this useful ... links in the comments.

46

6 Kommentare

📢 Our colleague, Michał Zyga, has written a great article explaining how JVM method profiling works. If you’re working with the JVM or simply curious about what’s happening under the hood, take a look - it’s well worth your time. 👉 Check out the article here: https://lnkd.in/dwakEJHd #JVM #Java #Profiling

5

Most royalty processing projects optimize for speed. The best ones optimize for recoverability. There's a difference — and it costs you later if you get it wrong. I've spent years building royalty and publishing infrastructure for mid-sized rights organizations. And the single most expensive mistake I see isn't bad software or slow teams. It's discarding the originals. Audio artifacts. Source recordings. Raw scan files. The things that feel redundant once the data is extracted. They're not redundant. They're your fallback layer. Here's the pattern I see, repeatedly: → Company processes a large catalog → Keeps only the extracted metadata → 12–24 months later, edge cases emerge (unknown works, ISRC mismatches, disputed recordings) → There's nothing to go back to → Manual reconstruction begins. Estimations. Disputes. We had this exact situation with a Swiss music rights cooperative — 200,000+ recordings annually. We kept the artifacts. We trained our AI models against them. We built the reconciliation layer on top of verified source material. When unknown recordings surfaced later? We reprocessed. We corrected. We delivered 95% metadata accuracy. Not because we were lucky. Because the architecture assumed that something would go wrong — and made recovery possible. Good infrastructure isn't about avoiding failure. It's about making failure recoverable. That's the standard I build to.

3

In our "Apache Airflow Best Practices for GenAI Pipelines" session, Tamara Janina Fingerlin points to some key Airflow basics to remember. https://bit.ly/4ngDbX0 #BeyondAnalytics

28

How to fix langgraph blocking code issue? If you would have productionized any langgraph agent beyond doing POCs/demos, then you would have experienced this for sure. This typically involves replacing synchronous I/O operations with async/await patterns, initializing models in the global scope rather than inside nodes, or running blocking code in separate threads using asyncio.to_thread for blocking functions Some Fix Scenarios: - Move Initialization to Global Scope: Initialize models (like ChatGoogleGenerativeAI) outside the node functions. Doing it inside nodes triggers synchronous authentication checks, causing blocks. - Use async/await Patterns: Ensure all LLM calls and tool calls are asynchronous. Switch from requests.get() to httpx.AsyncClient or aiohttp. - Run Sync Code in Threads: If you must use a blocking, synchronous library, wrap it with await asyncio.to_thread(sync_func). - Development Quick Fix: Run your dev server with {Link: langgraph dev --allow-blocking. - Deployment Fix: Set the environment variable BG_JOB_ISOLATED_LOOPS=true to handle blocking calls safely, although this can be slower. - Update Dependencies: If using Langfuse for tracing, ensure you are using version 3.3.3 or later to fix asynchronous context propagation issues.

1

Kafka's "noisy neighbors" problem (where workloads often compete for resources) can lead to increased latency, performance degradation, and failures. We review how to handle it with quotas, mirroring, and WarpStream Agent Groups ➡️ https://lnkd.in/eX5qH-Wm

10

Set up a modern lakehouse with Apache Doris + Gravitino Architecture: - Apache Doris: Acts as a real-time query engine, on top of S3 and Iceberg. It queries data across structured, semi-structured (JSON), and vector data, including Iceberg tables stored in S3, with sub-second latency. - Gravitino: Serves as the unified metadata and security layer. It connects Doris to Iceberg through a single REST interface and securely manages credentials via dynamic credential vending. - Iceberg: Provides the open table format that organizes data files on S3 into structured, queryable tables. It ensures reliable schema evolution, partitioning, and data consistency for the lakehouse. - S3: Acts as the underlying object storage layer, storing Iceberg table data in open formats (Parquet/ORC): cheap, scalable, and cloud-native. ➡️ Read the blog below to get a step-by-step guide: https://lnkd.in/gCgGyQQa #lakehouse, #ApacheDoris, #Gravitino, Datastrato, #S3

13

1 Kommentar

Are you looking for a drama-free way to run your Airflow pipelines? In a new example commented below, we show how 📦Kubetorch🔥 cleanly separates the concerns of development from orchestration, so you never need to get stuck push-and-pray debugging in Airflow again. You can iterate and debug your application in local Python with Kubetorch's live dispatch to cloud hardware, and then simply run that Python code as-is in Airflow with PythonOperator in CI or production. The Airflow node will run your code identically, on identical cloud hardware, as you ran and debugged it locally, so you'll never need to rely on Airflow as your debugger again. If you need to debug something that's breaking in production, or simply experiment with a change to a production pipeline, you can just edit and run that same Python code locally. It's a totally closed research-to-production-to-research loop. This example demonstrates how to use Kubetorch to train a PyTorch model on a remote GPU, and then schedule that training for recurring automation with Airflow. If you already have a bunch of Airflow pipelines, you'll see that it takes minimal code changes to restore the development and debugging experience in any existing pipeline. It's not just Airflow; the same structure can be used to improve development velocity, research-to-production, and fault tolerance with any pipeline orchestrator (e.g. Argo, Dagster, Prefect, Flyte) without requiring any direct integration. Take a look and get in touch if you'd like to try it out yourself!

12

1 Kommentar

Still running Bitbucket Pipelines? If you’re planning a move to GitHub, your CI/CD workflows need more than a lift-and-shift. Our latest blog breaks down how to migrate Bitbucket Pipelines to GitHub Actions without breaking builds, losing logic, or slowing teams down. ⚡ Cleaner pipelines 🔐 Better security 🔄 More flexible automation Read the full guide 👉: https://lnkd.in/dUs9ypyh ✍️ Tyurkiyan Mehmedova, DevOps and Cloud Engineer, ITGix Ltd #DevOps #CICD #CI #CD #GitHubActions #Bitbucket #PipelineMigration #PlatformEngineering #CloudNative #Automation #SoftwareDelivery #EngineeringLeadership

16

Great read by Ben Lorica 罗瑞卡 on Substack - "Inside the race to build agent-native databases" - which includes a review of Tiger Data (creators of TimescaleDB)'s recent launch of Postgres for Agents: https://lnkd.in/gxiSr_pW Quoted from the article: Tiger Data’s “Postgres for Agents” takes an evolutionary, rather than revolutionary, approach. Instead of building a new database from scratch, it enhances PostgreSQL, the popular open-source database, with capabilities tailored for agents. The cornerstone of this initiative is a new storage layer that enables “zero-copy forking.” This allows a developer or an agent to create an instantaneous, isolated branch of a production database. Try out our new Tiger Data CLI, which makes it easy to explore, fork, and build with Agentic Postgres: https://lnkd.in/gPZ5d-xH

10