Rene Svendsen’s Post

In the evolving landscape of artificial intelligence and machine learning, the integration of MLOps has become paramount for organizations seeking efficiency and scalability in their projects. Best End-to-End Open Source MLOps leverages a variety of platforms, frameworks, and tools designed to streamline the deployment and management of machine learning models. This resource provides valuable insights into the most effective open-source solutions available, allowing practitioners to select the right tools for their specific requirements. For more information, visit:

🔶 The Operating System Your Organization Doesn’t Know It Needs #️⃣ Infrastructure Gets Smart: Why Topology Finally Matters ▶️ Cloud used to treat every GPU as equal. Now, as AI workloads explode, placement is everything. With models spanning 80GB+, distributed training and topology-aware scheduling (think Kueue, NVLink, RDMA) have moved from “nice to have” to essential. The infrastructure itself is learning to optimize—not just run. #️⃣Model Context Protocol (MCP): From Integration to Intelligence ▶️Major vendors are racing to build MCP-driven tools. Why? MCP turns isolated agents into semi-autonomous coworkers, transforming everything from data infrastructure to service ecosystems. It’s the nerve center for operational collaboration—making bank operations, payments, and compliance systems fluid and future-ready. #️⃣Context Engineering: The Real AI Discipline ▶️Forget prompt engineering; context engineering is the new game. It's about designing knowledge flows for LLMs—using minimal prompts, summarization for long tasks, and just-in-time retrieval. This discipline unlocks reliable AI for complex banking workflows and multi-step financial processes #️⃣AI Coding: Maturity With Caution ▶️Coding agents—Claude Code, Cursor, Windsurf—have evolved. But beware: AI-generated code is duplicating, not refactoring; and ungoverned shadow IT is growing. Embed compliance into pipelines, shift from single prompts to team-wide instructions, and use structured output contracts to stay ahead. #️⃣Data Engineering Patterns: Lakehouses vs. Knowledge Graphs ▶️Two futures are emerging—unified, real-time lakehouses for fast, fresh data (Paimon, StarRocks); and knowledge graphs for memory-rich, agentic systems (RelationalAI, Mem0). Traditional ETL teams are building old architectures in a new era. The winners: those who treat data products with domain ownership, not isolated silos. #️⃣Patterns Worth Embracing ▶️Arm compute as the new default: Multi-arch Docker offers agility. ▶️Continuous compliance: Don’t bolt it on—bake it in. ▶️Context engineering as a practice: Treat it like design, not improvisation. ▶️MCP servers as first-class citizens: Don’t settle for API conversions. ▶️On-device info retrieval for privacy-first workloads: RAG on the edge, not just the cloud. ▶️Service mesh, no sidecar: Istio ambient mode is here. #️⃣Antipatterns to Watch ▶️AI-accelerated shadow IT: No-code tools stitch ungoverned, risky apps. ▶️Direct API-to-MCP conversions: Creates inefficiency, wastes tokens. ▶️Text-to-SQL shortcuts: Promise much, deliver little; use semantic layers instead. These tools and patterns aren’t predictions—they’re reality. The question: Are you optimizing the present, or rearchitecting for the future? #AgenticAI #TechRadar #CloudNative #FinTech #BankingTransformation #AIOperations #Innovation #Architectingforfuture

𝐌𝐂𝐏 𝐡𝐢𝐭 1,000+ 𝐬𝐞𝐫𝐯𝐞𝐫𝐬 𝐛𝐮𝐭 𝐰𝐢𝐭𝐡𝐨𝐮𝐭 𝐫𝐞𝐫𝐚𝐧𝐤𝐢𝐧𝐠 𝐲𝐨𝐮'𝐫𝐞 𝐨𝐧𝐥𝐲 𝐠𝐞𝐭𝐭𝐢𝐧𝐠 𝐡𝐚𝐥𝐟 𝐭𝐡𝐞 𝐯𝐚𝐥𝐮𝐞 The MCP ecosystem blew past 1,000 servers by early 2025, connecting everything from Slack to databases to ERP systems, and over 90% of the AI industry now supports MCP as the standard for AI interoperability. But here's the bottleneck nobody talks about: 𝐫𝐞𝐭𝐫𝐢𝐞𝐯𝐚𝐥 𝐰𝐢𝐭𝐡𝐨𝐮𝐭 𝐫𝐞𝐫𝐚𝐧𝐤𝐢𝐧𝐠 is leaving money on the table. Reranking can improve retrieval quality by up to 48%, and companies using hybrid retrieval with reranking are reporting 25% reduced token usage and cost. The architecture is simple but brutal in effectiveness: your MCP client pulls context from multiple sources, dense vector search narrows it to 50 to 100 candidates, then cross encoder rerankers apply semantic scoring to surface what actually matters. Databricks is seeing 89% recall at top 10 results with reranking, a 15 point improvement over baseline, all delivered in latencies as low as 1.5 seconds over 50 documents. The shift isn't just technical, it's economic. MarketsandMarkets predicts the global MCP market will reach $1.8 billion by 2025, and the companies winning are the ones stacking MCP servers with intelligent retrieval pipelines that combine semantic embeddings, hybrid search, and adaptive reranking before generation. Conversational AI in 2025 doesn't just connect to tools. It connects intelligently, ranks ruthlessly, and generates precisely. #MCP #Reranking #ConversationalAI #RAG #SemanticSearch

🤖 AI Agents meet the Code Property Graph ⚙️ I came across this recent post on an MCP server for the open-source code analysis platform Joern. This seems like a compelling direction to take: natural language is converted into code analysis queries, lowering the entry barrier significantly. https://lnkd.in/dYe63gDt

🔥 The Future of APIs: From REST to GraphQL to AI-Powered Interfaces APIs have quietly become the backbone of the digital world — yet the way we design and consume them is changing faster than ever. We started with REST. Then came GraphQL. And now, we’re entering an era where AI is redefining what “interfaces” even mean. 💡 1️⃣ REST: The Standard That Scaled the Internet 🔹 REST brought structure, predictability, and scalability. 🔹 It allowed developers to expose and consume resources through simple HTTP operations. 🔹 But it came with limitations — over-fetching, under-fetching, and rigid endpoints. 💡 2️⃣ GraphQL: A Smarter, Client-Driven Revolution 🔹 GraphQL gave clients control over the data they needed — no more wasted payloads. 🔹 It improved performance, reduced bandwidth, and made integrations flexible. 🔹 Yet, it introduced complexity: caching, schema management, and versioning challenges. 💡 3️⃣ AI-Powered APIs: The Next Interface Frontier 🔹 Instead of querying data, imagine asking an API in natural language. 🔹 LLM-powered systems (like OpenAI’s API or Anthropic’s Claude API) no longer expose endpoints — they expose intelligence. 🔹 Developers move from designing “routes” to designing capabilities. 🔹 The interface itself adapts dynamically to the intent, not a fixed schema. ✅ What This Means for Developers: 🔹 API design is shifting from structural to semantic. 🔹 Documentation will evolve from “endpoints and examples” to behavioral contracts — describing what an AI can do, not just what it returns. 🔹 Testing and security will need new paradigms to handle non-deterministic responses. 💬 The Big Question: As APIs become intelligent, do we lose predictability — or gain a new level of flexibility? Will the next generation of developers even need to “call APIs” the traditional way? What do you think — are we entering the post-REST era of software development? #APIs #SoftwareEngineering #GraphQL #RESTAPI #ArtificialIntelligence #LLMs #SystemDesign #DeveloperTools #FutureOfTech #DevCommunity

TOON can significantly boost AI pipeline performance by shrinking prompt sizes, cutting compute load, and accelerating interactions across models and services. With a projected 30–70 percent reduction in tokens, it has the potential to meaningfully lower the cost of using LLMs. However, for deeply nested or non-uniform data, JSON may be more efficient. https://lnkd.in/gV8qiWSF

🚀 Just wrapped up my next article: Beyond APIs — Lessons from Building with the Model Context Protocol (MCP) MCP is one of those ideas that seems small at first — a way to share context between tools — but once you implement it, you realize it’s a shift in how systems, developers, and AI agents collaborate. In this piece, I explore: • Why MCP is gaining traction • How natural language becomes the new abstraction layer • Lessons learned from integrating MCP into c4o • Why APIs (with security, identity, and observability) still matter more than ever 💬 Curious to hear your thoughts: Do you see MCP as the next evolution beyond APIs, or just another layer in the stack? #MCP #AI #Automation #KnowledgeGraph #Choreo #SchemaDriven #APIs #Context #Kubenet

In my company there is a strict verification process for internal and external MCP servers before they show in our AI Hub. There are dozens of approved MCP's today and as developers we are tempted to configure as many as we find useful. Recently i was working on an MCP server to define some useful sprint tools including things such as a Standup Assistant and a Smart PR (Pull Request). A Smart PR is defined as an MCP Prompt Resource. It uses tools from dual MCP's one being GitHub MCP and the other which updates our Work Tracking System with the PR information in the Work Chatter Feed. So when i naively tried to configure the GitHub MCP server Pop came the warning from my AI Coding Agent about exceeding total tool limits. This was because the GitHub MCP server was exposing a whopping 87 tools. I wanted to understand how it impacts different aspects : context, cost, reliability of responses etc. Here is a result of my exploration again using AI's help. 🧠 Technical Deep Dive: Tool Limits & LLM Context Context Window Impact: Each MCP tool consumes context space in the LLM's prompt: Tool Definition Overhead: Per Tool: ~200-500 tokens for tool schema definition 87 GitHub Tools: ~17,400-43,500 tokens Token Usage Breakdown: Total Context Window: ~200,000 tokens (Claude 3.5 Sonnet) ├── System Prompt: ~5,000 tokens ├── Tool Definitions: ~20,000-45,000 tokens (with all MCP servers) ├── Conversation History: ~10,000-30,000 tokens ├── User Query: ~100-1,000 tokens └── Available for Response: ~120,000-160,000 tokens Performance Implications Tool Discovery Time: 87 Tools: ~2-5 seconds for tool enumeration 10 Tools: ~0.3-0.7 seconds for tool enumeration Impact: Slower response times, especially on first interaction Memory Usage: Tool Schemas: Stored in memory for each conversation Context Compression: LLM may compress tool definitions, reducing accuracy Token Efficiency: Fewer tools = more tokens available for actual work 💰 Cost Implications of Context Overload Token Pricing Impact: Claude 3.5 Sonnet: ~$3.00 per 1M input tokens, ~$15.00 per 1M output tokens Tool Definition Overhead: 20,000-45,000 tokens per conversation Cost Per Conversation: $0.06-$0.135 just for tool definitions Context Window Efficiency: High Tool Count: Only 65% of context available for actual work Impact of Optimized Tool Count: Better responses, fewer follow-up questions, reduced token usage. Hidden Costs: Retry Attempts: Context overload can cause failed tool calls, requiring retries Follow-up Questions: Incomplete responses due to context limits Tool Selection Errors: Too many tools can lead to wrong tool selection Response Quality: Compressed context may reduce response accuracy TLDR; Disable tools that you do not need upfront. PS: I generally write about my work on medium.com but figured medium site is not accessible since yesterday (AWS Outage Impact). So decided to open my post account here. You can find my Medium posts at https://lnkd.in/gS7ydd8x

#JSON vs #TOON — A new era of structured input? https://lnkd.in/g9tC-pki AI is becoming cheaper and more accessible, but larger context windows allow for larger data inputs as well. LLM tokens still cost money – and standard JSON is verbose and token-expensive: TOON conveys the same information with fewer tokens Repo : https://lnkd.in/g4upeS5D