Mathieu Bastian

Most ML teams don’t struggle with models, they struggle with everything around them. Environment setup takes days, Iteration cycles take weeks, Evaluation pipelines break silently, The model is rarely the problem. This feels solvable

16

Supercharge your RAG apps with clean, structured data, not guesswork. LLMs don’t fail because the model is weak. They fail because the data is messy. Octro takes PDFs, scanned docs, and complex tables and turns them into reliable, structured JSON/CSV ready for vector search, embeddings, or downstream analysis. No token limits No hallucinations caused by noisy text No manual cleaning or regex nightmares If you're building: • RAG pipelines • AI Knowledge Bases • Data Agents • Automated Workflows Then your model deserves clean input. Your output is only as smart as your data. Make it trustworthy. 👉 https://octro.io

1

Nobody brags about data validation. But every engineer prays for it. Because nothing kills momentum faster than chasing: Random NULLs in production API payloads that don’t match Customers that don’t exist 🚀 Next wednesday, we’ll show you how those checks can run themselves, so devs can finally ship.

1

Test writing is hard to automate. It’s an exponential search space with tangled dependencies, frameworks, I/O concerns, and no easy way to measure quality. At Diffblue, we tackle this using Reinforcement Learning. We treat the JVM as our environment: generate test candidates, measure coverage and results, then reward “better” tests and iterate. Each cycle improves accuracy, efficiency, and code readability. AI doesn’t just write tests—it learns how to write high‑quality ones on its own. 📈 Want smarter, faster unit testing? Learn how we power developer productivity by automating the entire cycle—from test creation to maintenance. Explore more: https://lnkd.in/gmEFJr34

8

Interested in using LLMs to rewrite poorly written queries? Read the ACM SIGMOD paper from our founder and one of his PhD students:

7

Our GenRewrite work is now officially accepted to appear at ACM SIGMOD 2026. Slow query performance is one of the most common complaints in database systems. Part of this is on query optimizers to do better, but a significant portion is caused by poorly written queries - whether from inexperienced users or auto-generated by modern BI tools like Looker, where it's not uncommon for queries to span hundreds of lines. When the gap between what the user wrote and what the optimizer needs is too large, traditional query optimization falls on its face. Whether it's a DBA trying to hand-optimize queries or applying rewrite rules, the problem remains. The traditional approach has been rewrite rules - pattern-matching transformations that replace query fragments with optimized equivalents. But these are rigid. A rule-based rewriter can only optimize patterns it already knows. If your query doesn't match an existing rule, you're out of luck. With LLMs now capable of complex reasoning, the natural question is: can we do better? Can LLMs discover rewrite opportunities that existing rules miss? This is what GenRewrite explores. The challenge is that simply prompting an LLM to "rewrite this query faster" doesn't work well on complex queries. Most outputs either fail to parse or produce incorrect results. GenRewrite introduces Natural Language Rewrite Rules (NLR2s) - human-readable descriptions that guide the LLM toward effective rewrites and enable knowledge transfer across queries. The system learns over time: as it successfully rewrites queries, it accumulates NLR2s and applies them to similar queries based on performance bottleneck analysis. Some examples of NLR2s that GenRewrite discovered: * "Replace the correlated subquery with a precomputed aggregate in a separate CTE" * "Use common table expressions (CTEs) to precompute aggregates" * "Replace a correlated subquery that computes an aggregate with an inline view that pre-aggregates data" We also use counterexample-guided iterative correction to fix semantic and syntactic errors, improving correctness rates significantly - 70% of GenRewrite's rewrites are semantically equivalent to the original, compared to 52% for baseline LLM approaches. On TPC-DS, GenRewrite achieves 2x or higher speedup on 25 queries - 35% more queries than other LLM-based approaches and 160% more queries than LLM-enhanced rule-based systems. Some individual rewrites hit 9x to 24x speedups. On SQLStorm-generated queries that the LLM never saw during training, the gap widens further. Kudos to my PhD student Jie Liu for his great work on this! Full paper: https://lnkd.in/eHBdvw7c #LLM #GenAI #GenRewrite #QueryRewriting #SIGMOD2026

16

1 commentaire

KinD clusters rarely survive longer than a demo. Every setup change means rebuilding from scratch. No real LoadBalancer without MetalLB. Images need to be pushed before they're usable. And collaboration? Not part of the design. vCluster in Docker (vind) changes this. Three commands. Less than three minutes. No homelab. No cloud costs. A real Kubernetes control plane in Docker, with built-in LoadBalancer, pause/resume, and secure cluster sharing. Invite a teammate into your actual running local environment. Or attach an external VM as a worker node over Tailscale VPN with a single curl command. When you're ready for production, move the same config to Helm and deploy on a real cluster. Like a Lego brick. Read the full breakdown by Artem Lajko ☁️ ⎈: https://lnkd.in/gcBK6VJi

48

1 commentaire

There are generally two ways to build Scala APIs that have a specification: - the code-first approach and - the API-first approach (schema-first) In the "code-first" approach you start with the coding right away. Schedules are tight, the pressure is real. The specification (if any) is mostly an afterthought. Consider annotations in Spring framework, you can always sprinkle some `@ApiParam` here and there in your controllers. And voila, you got yourself an OpenApi spec generated from code In the "API-first" approach you start with the API specification. You define the API usually as an OpenAPI (Swagger) spec in a YAML or JSON format. Then you generate the server from that spec (and client code etc). We've always been advocates for the API-first approach - but how exactly do you do this in Scala? Check out our latest blog post to learn how 🔗 👇

5

4 commentaires

What if you could just ask your Kafka topic a question and get the answer in plain language? Many orgs rely on Apache Kafka® for real-time data, but not everyone is fluent in Flink SQL. That’s where AI assistants come in. In this demo, see how tools like Cursor and the Model Control Protocol (MCP) make it possible to ask natural language questions and get real answers directly from your Kafka topics on Confluent Cloud. Ready to try it now? Sign up for Confluent Cloud. ➡️ https://cnfl.io/419WTuH

37

1 commentaire

We treat metadata as a strategic layer because it's the connective tissue that powers everything from AI observability to data product discovery. Why? Because it's how we: - track lineage for RAG pipelines. - scale access without chaos. - debug LLM hallucinations. - earn stakeholder trust when GenAI goes live. The old model treated metadata as a compliance afterthought. The new model? It’s your leverage point. That’s why we’re doubling down on open standards (like OpenMetadata), platform-native catalogs (like Unity), and metadata-aware architectures that don’t just document — they activate. If you’re building anything with AI or advanced analytics, metadata is the foundational infrastructure.

15

Airflow orchestration is not a prettier cron. It’s the contract for how your data stays reliable. We inherited a shop with 40 noisy DAGs, blocking sensors, and retries hiding bad data. Here’s what turned it around: - Model dependencies with Datasets/clear trigger rules; kill cross-DAG sensors. - Make every task idempotent and add data checks before writes; fail fast. - Let Airflow conduct, not perform: push heavy compute to your engine; use deferrable operators and SLAs for “stale” alerts. Incidents dropped, backfills became boring, and on-call sleep returned. How do you decide what belongs in Airflow vs in your transformation layer? #Airflow #DataEngineering

6

Your ML models aren’t static, they’re living systems. And when they drift, the impact isn’t small. A 1% precision drop can cascade into major business risk. That’s why real-time AI observability isn’t a “nice-to-have.” It’s the safety net that keeps your models healthy, your compliance intact, and your business moving fast. Learn: -Why traditional monitoring fails in production -How real-time observability catches drift before it snowballs -The infrastructure you need to scale AI responsibly   If your agents are live, observability isn’t optional, it’s survival. Read More : https://hubs.li/Q03Stwhl0

14

AI agents are only as powerful as their memory. Redis has been named the #1 data storage choice for AI agents in 2025. The reason? Redis delivers what AI truly needs: - Real-time vector storage - Semantic caching - Long-term memory for LLMs At Bassirah, we see this as validation of why Redis is central to the next generation of AI-native data platforms powering everything from agentic RAG systems to enterprise decision engines. 🔗 Read more: https://lnkd.in/db47k6zG #AI #Redis #RAG #LLM #DataEngineering #AIAgents #ArtificialIntelligence

2

"I'd love to connect Claude Desktop directly to the database I'm working on for that contextual understanding." We asked TigerData engineer John Pruitt about MCP in workflows. His focus: Direct database context integration to make development environments more intelligent and context-aware. MCP deep–dive article: https://lnkd.in/gUk_WnGW

4