Leo Paillier

💡 LLMs have trouble with nested sequence of API calls with parameters that depend on response from previous calls. Can MCP help? For example, when the LLM generates a sequenece of API calls to sort out a problem: "What is the difference between the squares of 4 and 3?" Calls needed to solve this: {"name": "square", "arguments": {arg_0: 4}, "label": "$var_1"} {"name": "square", "arguments": {arg_0: 3}, "label": "$var_2"} {"name": "subtraction", "arguments": {arg_0: var1.resultvar1​.result, arg_1: var2.resultvar2​.result}, "label": "var_3"} The best performing models can get through these tests successfully only 25% of the time! Crazy! Connecting these APIs to the LLM via MCP by itself doesn't directly solve this problem, but it can help. You can think of the MCP server as a smart, stateful orchestrator of the backend API calls, and can do things like store the results between different tool calls and let the LLM refer to the results by variable name (resolving the actual values behind the scenes). This is an example of a stateful MCP server, and the spec supports this. I thought this was a great paper I stumbled upon from a post by Paul Lundin "NESTFUL: A Benchmark for Evaluating LLMs on Nested Sequences of API Calls" https://lnkd.in/d369aXgR

22

1 Comment

Most teams don’t fail at Iceberg because of the tech — they fail because of everything around it. You can have the best format, fastest engine, and cleanest tables… and still sink your Lakehouse if your tactics and strategy don’t align. On October 16, Yingjun Wu and I will talk about why Iceberg adoption can be tricky — and how not to mess it up. Expect real stories, tactical advice, and maybe one or two jokes about data teams turning schema evolution into a full-time job. 🧊 This isn’t just about config files — it’s about mindset, ownership, and the reality of running open data infra in production. 🎟️ Save your seat: https://luma.com/fga9usb8 #Lakehouse #Iceberg Apache Iceberg #DataEngineering Vakamo RisingWave

17

1 Comment

Orchestration is Dead Weight for AI Kubernetes. Containers. Scaling scripts. Heavy, reactive, and burning money. Cold starts, glue code, YAML churn, dashboards on dashboards... just to keep things running. At TAHO, we’ve built the end to all of it. TAHO is faster, lighter, and scales without the baggage. Orchestration had its moment. Now it’s time to move on. #PostCloud #AIInfra #TAHO #AIEfficiency

27

5 Comments

Elasticsearch has no native support for materialized views. No product on the market offers SQL-defined, continuously refreshed materialized views for Elasticsearch — until now. "How do you JOIN data across Elasticsearch indices?" The answer, for most teams, looks like this: 1. Build a Kafka consumer 2. Listen for changes in index A 3. For each change, look up data in index B 4. Write enriched documents to index C 5. Build another consumer for changes in index B 6. Re-enrich everything when B changes 7. Handle race conditions, retries, backpressure 8. Monitor, alert, maintain... forever Result: 2,000+ lines of code. 3+ Kafka topics (often 5-6 in production). An on-call rotation just for the denormalization pipeline. Now imagine this instead: CREATE MATERIALIZED VIEW orders_with_customers_mv REFRESH EVERY 10 SECONDS AS SELECT o.id, o.amount, c.name AS customer_name, c.city AS customer_city FROM orders AS o JOIN customers AS c ON o.customer_id = c.id WHERE o.status = 'completed'; One SQL statement. That's it. No additional infrastructure. No consumers. No application code. Everything runs on your existing Elasticsearch cluster. Under the hood, the engine orchestrates: → Elasticsearch transforms for continuous sync → Enrich policies for the JOIN → Ingest pipelines for computed columns → Automatic rollback on deployment failure Then query it like a regular table: SELECT * FROM orders_with_customers_mv WHERE customer_city = 'Paris'; Sub-millisecond response. Pre-computed. Continuously refreshed. Want aggregations too? CREATE MATERIALIZED VIEW revenue_by_city_mv AS SELECT c.city, SUM(o.amount) AS revenue, COUNT(*) AS orders FROM orders o JOIN customers c ON o.customer_id = c.id GROUP BY c.city; Real JOINs in Elasticsearch. From SQL. No application code. Note: automatic enrich policy re-execution (when lookup data changes) requires an Elasticsearch Platinum license (Watchers). Without one, schedule REFRESH MATERIALIZED VIEW yourself — a cron job, a Kubernetes CronJob, or an Airflow task. Transforms, enrich policies, and REFRESH EVERY work on any free cluster (ES 7.5+). Available in the community edition (up to 3 views). Full documentation: https://lnkd.in/eBXzUmQZ 💻 GitHub: https://lnkd.in/emxmZwtE 🌐 Website: https://softclient4es.dev/ How does your team handle cross-index data in Elasticsearch today? #Elasticsearch #MaterializedViews #DataEngineering #SQL #JOINs #Architecture

3

1 Comment

I am at Astronomer because I know that 1) ML and AI both start and end with data, 2) the data pipelines for these challenges are exploding exponentially in complexity, and 3) great orchestration is the only way to stay ahead of that curve. Airflow and Astronomer is already solving this for big companies you rely on every day. OpenAI runs over 7,000 unique complex data worfklows in Airflow to make their magic happen (yes, that's a lot), and they expect that to grow 10x. They built a lot of customizations into their private fork of Airflow that already exists natively in Astronomer's. If you have ambitions for AI or ML in your company, start with your orchestration strategy.

44

1 Comment

Thrilled to support AWS Bedrock TOOL PROMPT CACHING on LiteLLM (YC W23) - this enables cutting costs for assistant tool responses IN THE SAME FORMAT as Anthropic API / Bedrock / VERTEX AI - https://lnkd.in/dW9sJDMS ✅ Mistral - translate new 'thinking' content block 🛠️ Mistral - filter out empty assistant content (causes API errors) ⚡️ Image Edits - support OpenAI 'mask' parameter 🎉 Responses - support 'allowed_openai_params'

10

How Attentive uses Gatling and Datadog to achieve 360° observability at scale For teams operating at scale, performance issues are rarely clear-cut. The real challenge is understanding whether traffic is flowing, where it slows down, and why. In an upcoming live session, Bian Jiang, Principal Software Engineer on Attentive's Developer Experience team, will walk through how Attentive connects Gatling and Datadog to gain end-to-end visibility during high-scale load testing. We’ll explore: - How Attentive correlates generated load with real system behavior - Why observability is critical to trusting load test results - Lessons learned testing systems under real constraints - How Datadog and Gatling together close blind spots at scale 📅 January 22, 2026 - 6PM CET / 12PM EST 🎙 Bian Jiang, Principal Software Engineer, Attentive 🗃️ Diego Salinas, Enterprise Content Manager, Gatling 🗃️ Heather Thacker, Developer Relations, Gatling 🔗 Register here: https://lnkd.in/eqvEuMhB

18

2 Comments

LLMs are stateless. They forget everything. That's the problem no one talks about. Google's Context Engineering paper exposed what production teams learned the hard way - memory architecture determines agent survival. ⚡ 𝗧𝗛𝗘 𝗦𝗛𝗜𝗙𝗧: 🧠 Sessions = Working memory (one conversation) 💾 Memory = Persistent storage (across conversations) ❌ Stateless → ✅ Stateful  ❌ Dumb chatbot → ✅ Intelligent agent Skip this. Stay irrelevant. ━━━━━━━━━━━━━━━ 🎯 𝗧𝗛𝗘 𝟯-𝗟𝗔𝗬𝗘𝗥 𝗦𝗧𝗔𝗖𝗞: 𝟭. 𝗘𝘅𝘁𝗿𝗮𝗰𝘁 → Knows you hate verbose answers 𝟮. 𝗖𝗼𝗻𝘀𝗼𝗹𝗶𝗱𝗮𝘁𝗲 → Updates when you switch from Python to Rust   𝟯. 𝗥𝗲𝘁𝗿𝗶𝗲𝘃𝗲 → Stops asking "What's your name?" in message #5 ━━━━━━━━━━━━━━━ 💡 𝗧𝗛𝗘 𝗜𝗡𝗦𝗜𝗚𝗛𝗧: Memory ≠ RAG database 📚 RAG = research librarian 🤝 Memory = personal assistant RAG finds facts. Memory learns YOU. ━━━━━━━━━━━━━━━ 🔴 𝗧𝗛𝗘 𝗥𝗘𝗔𝗟𝗜𝗧𝗬: Your agent just asked for the account number. Third time today. The $500 API bill hurt.  The 1-star review hurt more. __________ ♻️ Repost to help your network stop building goldfish AI ➕ Follow Mohammad Syed for more insights on AI and Cybersecurity.

61

57 Comments

🚀 AI Minute Mondays: Data Uptime & Availability for AI with Mandar Naik, Global VP at Redis. This week, I had the chance to sit down with Mandar Naik, Global VP of Sales and Partnerships at Redis to unpack one of the most underestimated pillars of AI systems: data availability. As AI workloads scale, the conversation can’t just be about models and GPUs. If your data isn’t available, consistent, and resilient, your AI simply won’t perform. Mandar and I dug into what “availability” really means in modern AI architectures—and why it’s far more than a checkbox.   We explored: 🔹 Redundancy as a design principle, not an afterthought 🔹 How latency and throughput shape real‑time AI performance 🔹 The role of failover and resilience in keeping systems running 🔹 Why consistency models matter more than most teams realize 🔹 The growing importance of monitoring & observability for AI data pipelines   We also talked about the common pitfalls teams run into when building for availability: ⚠️ Not planning for scale early enough ⚠️ Treating availability as a “nice to have” ⚠️ Relying on a single data source ⚠️ Failing to design for failure (because failure will happen)   Mandar brought clarity to a topic that every AI‑driven organization needs to take seriously. If you’re building AI systems that need to be fast, reliable, and production‑ready, this conversation is worth your time. Stay tuned for the full episode—and let me know what topics you’d like us to explore next on AI Minute Mondays. #AIMinuteMondays #AIMinuteswithSuchi #DataUptime #DataAvailabilityforAI #AILearning

61

1 Comment

Semantic layers are being built by infrastructure people for infrastructure people 🔧 ... and it shows. Every vendor loves to demo their sophisticated data modeling capabilities. Look at all these YAML configurations! Check out these complex relationship definitions! Marvel at our advanced metric inheritance patterns! Meanwhile, business users take one look at the interface and nope the f*ck out 🏃‍♀️ The UX is always an afterthought because the people building semantic layers usually have never actually had to use them day to day. They're optimizing for technical elegance, not for the poor marketing manager who just wants to know why their campaign numbers don't match last week's report. You end up with incredibly powerful systems that nobody adopts because learning the interface feels like getting a second degree in data architecture 📚 This is why semantic layer implementations have such brutal failure rates. It's not that the technology doesn't work. It's that the experience is designed for the data team, not the humans who actually need the insights. If you want to build a semantic layer that people will actually use, you need to start with the user experience and work backwards. Not the other way around. At Lightdash we built the BI tool first and when you watch real users struggle with data exploration every day, you learn pretty quickly what makes interfaces intuitive vs what makes people want to throw their laptop out the window 💻➡️🪟 The best semantic layer is the one that disappears completely. Users shouldn't even know it exists. Sound familiar or have you had a rip-roaring success with your SL?

48

19 Comments