Mourad Askar

Ever feel like today’s AI models are getting smarter but somehow less useful? You’re not imagining it. LLM benchmarks have become the obsession. Model makers are locked in a cycle of chasing leaderboard points on MMLU, GSM8K, and other tests no one actually cares about outside of research labs. The result? Models that ace exams but fall apart in real-world use. They misinterpret vague instructions. They forget what you said three messages ago. They can’t think beyond a single step. But sure, the score went up by 0.3 percent. Benchmarks were supposed to be a guide. Instead, they’ve become a crutch. A way to show “progress” without improving anything that matters. The questions we should be asking are simple: Can the model follow messy, human instructions? Can it reason through problems, not just regurgitate answers? Can it actually help people get real work done? If you’ve worked with these models, you’ve felt the gap. You’ve seen the hype, and then you’ve seen it fall short when it counts. Are we building for people, or are we building for points? I’d love to hear what others are seeing in the wild. #AI #LLMs #Benchmarks #MachineLearning #RealWorldAI

27

3 Comments

Single-method retrieval often fails in #financial and #regulatory domains, exact-term search misses semantic relationships, while dense models overlook structured identifiers. A multi-path approach resolves this gap. The Multi-Path Retrieval (#MPR) pipeline implements four complementary retrieval strategies with a domain-specialized re-ranking mechanism to balance lexical precision and semantic understanding: #BM25 Sparse Retriever – captures exact matches for regulatory codes, ticker symbols, and key phrases. #Dense Semantic Retriever – identifies paraphrases and conceptually similar expressions through high-dimensional embeddings. #Metadata-Aware Retriever – filters at section-level granularity, narrowing retrieval to relevant document parts before detailed chunk search. #HyDE Retriever – generates a hypothetical answer, embeds it, and retrieves real text segments with similar structure and terminology, bridging the question–answer linguistic gap. Final results can be aggregated through weighted rank fusion (like semantic > BM25 > metadata > HyDE ), producing top-ranked chunks enriched through context bundling when adjacent similarities exceed 0.85. This ensures coherent, context-complete passages are provided to downstream models. Such architecture is especially effective in financial compliance and regulatory analysis, where both formal precision and contextual interpretation are essential. The hybrid fusion enhances #recall, #contextualaccuracy, and #interpretability in retrieval-augmented systems(#RAG). #GenAI #Banking #AIEngineer #LLM #Fraud

4

AI is shifting from single models to ecosystems of specialized agents. This research on Reinforced Agent Merging (RAM) highlights a critical reality:  If you merge RL-trained agents the same way we merged fine-tuned models, you destroy value. Why it matters:  • RL agents learn sparse, task-specific intelligence  • Naive averaging dilutes the very behaviors you trained them for  • “One model to rule them all” is no longer viable What RAM shows instead:  • Share common intelligence  • Preserve and amplify unique capabilities  • Combined systems can outperform individual specialists Impact for leaders and builders:  • AI platforms will be agent-based, not model-centric  • Competitive advantage shifts to orchestration and integration  • The hardest problem is no longer training models, but merging intelligence without loss This is where AI strategy meets product leadership. If you are still designing around a single assistant, you are already behind. #AI #ArtificialIntelligence #AgenticAI #ReinforcementLearning #GenAI #AIArchitecture #AIProductManagement #AILeadership #EnterpriseAI https://lnkd.in/ed7pTrsD

7

Many “propensity logs” in production are actually proxies: a model score, a rank position, or a probability computed before downstream filtering and fallbacks. For Off-Policy Evaluation (OPE), the required quantity is different: the probability that the production system chose the action it chose, given the request context and the internal state it read at decision time. That probability needs to be recorded 'after' all business rules, gating, fallbacks, and any randomness. I wrote a Medium article that explains what “correct propensity” means in real systems, where it commonly breaks in production pipelines, and a minimal logging schema that still supports debugging and audits. Article: https://lnkd.in/eUwhCQhe #Experimentation #ABTesting #DataEngineering #MachineLearning #CausalInference

4

𝗤𝘂𝗲𝗿𝘆 𝗲𝘅𝗽𝗮𝗻𝘀𝗶𝗼𝗻 is one of the simplest yet most effective ways to improve retriever performance in RAG systems. While working on production-grade 𝗥𝗔𝗚 𝗽𝗶𝗽𝗲𝗹𝗶𝗻𝗲𝘀, I’ve consistently seen how expanding a single query into multiple variations significantly improves retrieval quality, especially 𝗿𝗲𝗰𝗮𝗹𝗹. In this article, I break down: What query expansion is Different techniques (𝗱𝗶𝗰𝘁𝗶𝗼𝗻𝗮𝗿𝘆, 𝗟𝗟𝗠, 𝗣𝗥𝗙) And a practical evaluation using 𝗣𝗿𝗲𝗰𝗶𝘀𝗶𝗼𝗻, 𝗥𝗲𝗰𝗮𝗹𝗹, and 𝗠𝗥𝗥 𝗥𝗲𝗮𝗱 𝗵𝗲𝗿𝗲: https://lnkd.in/d8KS6PUi If you find it useful, feel free to share it with others. :) #ArtificialIntelligence #MachineLearning #RAG #DataScience #LLM

4

Are AI-Powered Data Analysts Ready for Prime Time? Two Competing Approaches—Your Thoughts Needed As someone deeply involved in data science and AI research, I am writing an internal paper on the future of AI-powered data analysts. I would appreciate the opportunity to gather insights from my network. **The Problem: Hype vs. Reality**   In recent years, many tech leaders have promoted the notion that agentic AI systems could significantly replace traditional analyst and data scientist roles. The current development of LLMs does not (and likely will not) support this. Initially, LLMs were merely next-word predictors that were often unreliable and prone to hallucination. Despite using human feedback (RLHF) for improving coherence, these systems still engage in what philosopher Harry Frankfurt termed "bullshitting"—issuing confident statements that are not factually grounded. To address this,"Chain-of-Thought" (CoT) prompting was introduced, which guides models through reasoning step-by-step. Recent studies, including those from Apple on DeepSeek-R1, GPT-4o, and Claude Sonnet, indicate that models likely replicate chain-of-thought patterns from their training data instead of genuinely generalising problem-solving strategies. Neuroscience research suggests that different brain regions are responsible for language and reasoning, indicating that analytical thinking involves something other than just matching language patterns. **Two Approaches: Conversational AI vs. Modular Agentic Workflows**  Much of what LLMs accomplish in analytics relies on "training data poisoning," where they recycle solutions from textbooks, GitHub, or online forums rather than solving the problem.  1. **Conversational AI on Self-Driven BI:**   This practical approach builds on "self-driven business intelligence" (BI) systems. In this setup, analytical work is prebuilt and maintained by experts, with a library of reliable analytical modules and dashboards. The LLM primarily serves as a conversational interface, translating natural language queries from business users into structured requests that map onto these prebuilt assets. The chatbot helps users explore insights and create reports. 2. **Modular Agentic Workflows:**   This more robust method involves breaking down analytical tasks into smaller, manageable steps, each handled by specialised agents or tools. An agent framework can be crafted to tackle common tasks, such as causal analysis and customer segmentation. This method goes beyond traditional AutoML by automating not just model selection and tuning but also managing the entire analytical process step-by-step, from data preparation to interpretation. **Your Feedback Needed:**   Have you tried either approach in your analytics workflows? Which approach—end-to-end LLMs or modular agentic workflows—seems more feasible or practical in your context? Looking forward to your thoughts.   #AI #DataScience #LLM #BusinessIntelligence #AgenticAI #DataAnalysis

4

1 Comment

𝗧𝗵𝗲 𝗻𝗲𝘅𝘁 𝗳𝗿𝗼𝗻𝘁𝗶𝗲𝗿 𝗶𝘀 𝗔𝗴𝗲𝗻𝘁𝗶𝗰. Large Language Models are shifting from passive responders to 𝗮𝗴𝗲𝗻𝘁𝗶𝗰 𝘀𝘆𝘀𝘁𝗲𝗺𝘀 that plan, use tools, and self-correct. Once you have loops, the hard questions become operational:   • How do you 𝘁𝗲𝘀𝘁 𝗮𝗻𝗱 𝗺𝗼𝗻𝗶𝘁𝗼𝗿 iterative behavior?   • What does 𝘁𝗿𝗮𝗰𝗲𝗮𝗯𝗶𝗹𝗶𝘁𝘆 look like when control flow is probabilistic? My latest Medium article, 𝗧𝗵𝗲 𝗗𝗲𝗳𝗶𝗻𝗶𝘁𝗶𝘃𝗲 𝗥𝗔𝗚 𝗥𝗲𝗮𝗱𝗶𝗻𝗴 𝗟𝗶𝘀𝘁 𝗳𝗼𝗿 𝟮𝟬𝟮𝟲, curates the papers that matter for building and assessing these architectures:   • 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗥𝗔𝗚 (loops, policies, budgets)   • 𝗚𝗿𝗮𝗽𝗵𝗥𝗔𝗚 (grounding with stronger provenance)   • 𝗦𝗲𝗹𝗳-𝗰𝗼𝗿𝗿𝗲𝗰𝘁𝗶𝗼𝗻 (quality gates and continuous evaluation) Full reading list: https://lnkd.in/dKNpuhxp #ViewsMyOwn #AgenticAI #RAG #LLM #AIGovernance #GenerativeAI #GraphRAG

10

**Text Chunking Strategies: Elevating LLM Performance** Are your Large Language Models (LLMs) struggling with context, missing key information, or even hallucinating when processing extensive documents? The often-overlooked secret to unlocking their full potential lies in effective text chunking. Text chunking isn't just about splitting data; it's about intelligently preparing information to ensure your LLM receives the most relevant and coherent context. This is especially crucial for robust Retrieval Augmented Generation (RAG) applications and efficient LLM operations. Let's explore key strategies shaping today's AI landscape: * **Fixed Chunking:** The most straightforward approach, where text is divided into uniform segments of a predetermined size. Simple and quick, but can sometimes inadvertently split crucial contextual information. * **Recursive Chunking:** A more sophisticated method that breaks down text hierarchically (e.g., document > section > paragraph > sentence). This approach excels at preserving logical structure, ensuring related ideas stay together within a natural flow. * **Semantic Chunking:** Leveraging Natural Language Processing (NLP) to group text segments based on their meaning and thematic coherence. This is invaluable for applications demanding deep contextual understanding, as it ensures conceptually linked content is presented as a single unit. * **Adaptive Chunking:** The cutting edge, dynamically adjusting chunk size and strategy based on factors like content type, query complexity, LLM token limits, or even feedback loops. It often combines elements of the above to achieve optimal, use-case specific results. Choosing the right chunking strategy is a critical architectural decision, directly impacting your LLM's accuracy, relevance, and computational efficiency. It's not a 'one-size-fits-all' solution, but rather a strategic choice driven by your data's characteristics and your application's specific requirements. Which chunking strategy have you found most effective in your LLM projects, and what unique challenges have you encountered? Share your insights and experiences below! #LLM #NLP #RAG #TextProcessing #AI #DataScience #MachineLearning #GenerativeAI

6

Ever wonder how AI systems can seamlessly integrate with external tools and data sources? Enter the "Model Context Protocol" (MCP) from Anthropic—a game-changer for AI system integration. What it is: The Model Context Protocol establishes an open-source standard designed to bridge the gap between AI models and external systems. It aims to solve the problem of AI systems operating in silos, unable to effectively communicate with other tools or access real-time data. The core contribution here is a structured framework that makes it easier for AI models to dynamically access and utilize contextual information from a variety of external systems. How it works: At its heart, MCP features a modular architecture where AI components communicate through standardized APIs, allowing real-time data exchange and context sharing. Think of it as a universal translator for AI systems, employing JSON-based data formats and RESTful APIs to ensure efficient and scalable integration across diverse applications. This flexibility is key—models can adapt to different operational contexts without the need for extensive retraining, making them more versatile. Why it matters: From my perspective as an engineer, MCP is a significant step forward in overcoming interoperability challenges in AI deployment. This protocol enhances the adaptability of AI models, particularly in dynamic environments that demand real-time data, such as healthcare and finance. Initial results are promising, with a 30% boost in efficiency for multi-task learning scenarios and improved accuracy rates when models access real-time contextual data. References: - Paper: https://lnkd.in/gGt8NPAi - Blog: https://lnkd.in/gshhz6NV - News: https://lnkd.in/gTmqqFip Question: How could adopting a standardized protocol like MCP change the way your team integrates AI models with external systems? #AI #MCP #Integration #Interoperability #MachineLearning

6

Today I explored different chunking strategies for RAG systems, and it’s clear that choosing the right approach directly impacts retrieval quality. Here are the key strategies I understand, along with their trade-offs: 1. Character-based splitting ✔️ Pros: Simple to implement Works consistently across all text types Predictable chunk sizes ❌ Cons: Breaks context mid-sentence Poor semantic understanding Lower retrieval accuracy 2. Recursive character splitting ✔️ Pros: Preserves structure better than fixed splitting Flexible fallback mechanism (paragraph → sentence → word) Balanced chunk quality ❌ Cons: Still not fully semantic Requires tuning (chunk size, overlap) 3. Document-specific splitting ✔️ Pros: Leverages document structure (headings, sections, tables) Better contextual grouping Ideal for PDFs, markdown, structured docs ❌ Cons: Format-dependent Requires custom parsers per document type 4. Semantic chunking ✔️ Pros: Context-aware splitting Higher relevance in retrieval Improves embedding quality ❌ Cons: Computationally expensive Requires embeddings during preprocessing Slower pipeline 5. Agentic / intelligent chunking ✔️ Pros: Dynamic and context-driven Can adapt per use case Potentially best retrieval performance ❌ Cons: Complex to design Higher cost and latency Harder to debug One approach I’m currently exploring as well and that is Unstructured.io. It looks promising for handling multi-format documents (PDF, HTML, DOCX, etc.) and extracting structured elements before chunking. I am Curious about what chunking strategy has worked best in your RAG pipelines and why ? #AI #RAG #Backend #DataEngineering #LLM #SystemDesign

5

Transformers in Robotics I was reading some of the recent advancements in robotics and observed that robotics is quietly borrowing architectures from NLP world. Not the buzzword stuff-real architectural shift. Transformers changed how we think about language, then vision, and now they're creeping into control systems. What caught my eye isn't the tech itself. It is the intent...stop building one robot brain per task. Instead, train a single model on messy, multi-modal trajectories - vision, proprioception, language, action and let it generalise. The same way a language model switches from translation to coding to summarising with ease. The upside is obvious, fewer brittle pipelines, less custom engineering, adaptation without starting from zero. The messy part? Data. Real-world robot data is expensive, biased, risky, and painfully slow to gather. Safety isn't a feature, it's a constraint. Latency isn't a metric, it's survival. I think the real battle isn't in clever architectures, it is in building the data engines, safety nets and deployment loops that don't fall apart outside the lab. We are inching toward - pretrain once, adapt everywhere. But robotics is not text. When a token misfires on paper, you edit. When it misfires on a robot, something breaks. #AI #Robotics

169

9 Comments

AI is reshaping how businesses operate, transforming roles, workflows, and data standards much more. As AI fluency grows, so does the focus on ethical, responsible AI and faster prototype-to-production cycles. One small topic to consider when designing heavy AI applications is file formats like TOON (Token-Oriented Object Notation). Think of it as a translation layer: use JSON programmatically, and encode it as TOON for LLM input in AI-powered apps. Check it out here: https://lnkd.in/d__dmssE TOON is production-ready, but also an idea in progress. I’d love to hear about experiences once others start using this or similar approaches.

4

A 10-second question that can instantly eliminate ML/AI "experts” in interviews. My quick screening question: “Have you ever seen a negative Gini in a classification model?” Silence - Confusion - Or a very interesting conversation. Anyone who has actually built and evaluated models in the real world has probably encountered this at least once — especially early in their career. Because in theory, Gini shouldn’t be negative. But in practice, it can happen. Usually when: The higher predicted probability corresponds to lower event rate The decile/bands are sorted incorrectly Someone forgot to sort probabilities in descending order before computing the lift/Gini Classic beginner moment. We’ve all been there. The model suddenly looks worse than random. And then you realize… “Oh wait — the bands are sorted the wrong way.” 😅 It’s a small detail, but it reveals something important: 👉 Whether someone has actually built, validated, and debugged models or 👉 only read about them in theory These tiny practical questions often reveal more than complex theoretical ones. This is my #4th post in the series: “Quick MLAI interview questions to save the interviewer time.” My question to the community:Have you ever encountered negative Gini in a classification model? What caused it in your case? Let’s hear your stories 👇 #MLAI #MLAIinterview

28

9 Comments

𝐑𝐀𝐆 𝐈𝐬 𝐄𝐯𝐨𝐥𝐯𝐢𝐧𝐠 – 𝐑𝐀𝐆 2.0, 𝐀𝐠𝐞𝐧𝐭𝐬, 𝐚𝐧𝐝 𝐋𝐚𝐧𝐠𝐆𝐫𝐚𝐩𝐡: 𝐖𝐡𝐚𝐭 𝐘𝐨𝐮 𝐍𝐞𝐞𝐝 𝐭𝐨 𝐊𝐧𝐨𝐰 🚀 The RAG (Retrieval-Augmented Generation) landscape is transforming rapidly. What started as simple "retrieve and generate" has evolved into sophisticated multi-agent systems that can reason, plan, and execute complex workflows. 𝐓𝐡𝐞 𝐄𝐯𝐨𝐥𝐮𝐭𝐢𝐨𝐧 𝐟𝐫𝐨𝐦 𝐑𝐀𝐆 1.0 𝐭𝐨 𝐑𝐀𝐆 2.0 𝐑𝐀𝐆 1.0 was straightforward:  → Embed documents → Retrieve similar chunks → Generate response 𝐑𝐀𝐆 2.0 is intelligent: → Multi-step reasoning → Query decomposition & refinement → Self-correcting retrieval → Context-aware generation 𝐖𝐡𝐚𝐭'𝐬 𝐃𝐫𝐢𝐯𝐢𝐧𝐠 𝐓𝐡𝐢𝐬 𝐄𝐯𝐨𝐥𝐮𝐭𝐢𝐨𝐧? 🔍 𝐒𝐦𝐚𝐫𝐭𝐞𝐫 𝐑𝐞𝐭𝐫𝐢𝐞𝐯𝐚𝐥 •Hybrid search (vector + keyword + graph) •Query rewriting and expansion •Contextual re-ranking •Multi-hop reasoning across documents 🤖 𝐀𝐠𝐞𝐧𝐭-𝐁𝐚𝐬𝐞𝐝 𝐖𝐨𝐫𝐤𝐟𝐥𝐨𝐰𝐬 •Planning and execution phases •Tool selection and usage •Memory management •Error handling and recovery 📊 𝐋𝐚𝐧𝐠𝐆𝐫𝐚𝐩𝐡 Revolution LangChain's LangGraph is the game-changer here. It enables: •Visual workflow design •Conditional logic and branching •State management across steps •Human-in-the-loop capabilities 𝐑𝐞𝐚𝐥-𝐖𝐨𝐫𝐥𝐝 𝐈𝐦𝐩𝐚𝐜𝐭: I've seen RAG 2.0 systems: ✅ Reduce hallucinations by 40-60% ✅ Handle complex multi-document queries ✅ Provide explainable reasoning paths ✅ Adapt to domain-specific requirements 𝐖𝐡𝐚𝐭 𝐓𝐡𝐢𝐬 𝐌𝐞𝐚𝐧𝐬 𝐟𝐨𝐫 𝐘𝐨𝐮 If you're building AI applications: 🎯 Start experimenting with LangGraph for complex workflows 🎯 Invest in better retrieval strategies beyond simple similarity 🎯 Design for explainability - users need to trust AI reasoning 🎯 Plan for scale - agent systems require robust infrastructure The future of enterprise AI isn't just about better models - it's about smarter systems that can reason, retrieve, and respond like human experts. 💡 What's your experience with RAG evolution? Are you seeing similar improvements in your AI implementations? #RAG #LangGraph #LangChain #AI #MachineLearning #Enterprise #TechLeadership #Innovation

7

Major microstructure changes and rough dates that you should know when backtesting with CME market data: Jan 2, 2013 - Messaging Efficiency Program becomes effective. Jun 3, 2013 - Self-match prevention functionality launched. Mar 9, 2014 - New iLink and MSGW architecture (Phase 1) launched for TAS markets. Sep 28, 2014 - CME introduces MDP 3.0 based on SBE. Sep 21, 2015 - BTIC launched. Jan 31, 2016 - New iLink and MSGW architecture pre-launch for remaining markets (Phase 2). Apr 10, 2016 - New iLink and MSGW architecture in production for remaining markets (Phase 2). Jul 10, 2016 - Major performance enhancements to CME Globex. (?) Jul 17, 2016 - Production rollout of iLink nanosecond timestamp enhancement and tag 5979. May 21, 2017 - Production rollout of MBO for final market segment. Jul 17, 2017 - Russell index futures and options return to CME. Sep 24, 2017 - Implied functionality for FX futures and SD calendar spreads. Dec 18, 2017 - CME launches crypto futures with Bitcoin futures. Mar 25, 2018 - FX Link launched. May 13, 2018 - TACO launched. Jul 29, 2018 - iLink performance enhancements. Jan 13, 2019 - 9-digit price precision extension. Feb 3, 2019 - Launch of consolidated iLink fill messages. May 6, 2019 - Micro E-mini futures launched. Jun 17, 2020 - Production rollout of binary iLink 3.0 for final market segment. Jun 27, 2021 - Major performance enhancements to CME Globex. Jun 27, 2021 - 3:15 - 3:30 PM CT pause eliminated for CME equity index futures and options. Sep 17, 2021 - Standard-sized S&P 500 index futures and options delisted. Apr 14, 2023 - Eurodollar futures and options converted to SOFR contracts. Jun 26, 2023 - Eurodollar futures and options delisted with expiration of June contracts. Jan 7, 2024 - Major performance enhancements to CME Globex. Jan 7, 2024 - Mass quote messaging capacity increased. Mar 25, 2024 - Micro Treasury futures launched. Need a microstructure-aware vendor and market depth going back to 2009? You get both with #Databento.

164

8 Comments