D V.’s Post

🌍 Can #AI Use All My Documents at Once? 👉 Missed earlier posts? Start with the intro https://lnkd.in/daxm_2AV or last week’s post: Are AI Citations Reliable?https://lnkd.in/dHDzn7cG Executives often assume that connecting an AI system to a company’s full document base makes it instantly more powerful. The reality is more nuanced. Large language models (LLMs) are not designed to process all available data simultaneously. They work within fixed context windows, a limited space in which relevant information is loaded during a conversation or query. Adding more documents doesn’t automatically make an AI-powered system smarter. Beyond a certain threshold, scale can reduce performance, increase computational cost, and blur the distinction between relevant and outdated information. The goal is precision, not exhaustion from constant cost controls. ▶️ Why this matters for mobility and tolling In regulated mobility domains such as #CITS, #V2X, and #tolling, completeness must never come at the expense of control. Not all documents are equal. And, not all docs are meant to be part of every decision. Architecture reviews often show how retrieval pipelines that pull “everything” can quietly undermine compliance: 🔸 outdated specifications remain accessible, 🔸 local tariff or toll domain updates overlap with expired ones, 🔸 delays in implementation of EU regulations throw guidance out of sync, 🔸 version control between source systems breaks down. The result: AI systems referencing redundant or conflicting material, producing outcomes that are technically correct but operationally invalid. Regulated infrastructure relies on curated knowledge, not unrestricted ingestion. Governed retrieval architectures define what is available and retrievable - when and under what authorization. Key safeguards include: 🔸 Access control that enforces data relevance and document scope 🔸 Metadata tagging for validity periods and jurisdiction 🔸 Retrieval thresholds that limit input size and prevent context overflow 🔸 Validation layers that flag conflicts between versions A compliant system must demonstrate that every data point used in reasoning was current, authorized, and relevant to its operational context. ⚡ Leadership takeaway More data does not equal more intelligence. In fact, unchecked access often reduces both performance and trust. Effective AI governance starts by defining boundaries: what the system can use, under what rules, and when updates are applied. The principle applies across industries: in tolling, financial services, and regulatory AI alike. Relevance is a compliance requirement, not an option. Leaders should prioritize structured curation over scale. Foresight comes from ensuring it uses only what matters, not just feeding AI everything. 👉 Next Wednesday: Can AI Hallucinations Be Eliminated? #Leadership #Governance #Trust #Compliance #Mobility

🌍 Are #AI Citations Reliable? 👉 Missed earlier posts? Start with the intro https://lnkd.in/daxm_2AV or last week’s post: Does AI Use Only My Documents? https://lnkd.in/d-WT7nzB Executives often assume that when an AI system presents a citation or a source, it can be trusted. In reality, large language models (#LLMs) do not verify references in the same way humans or structured systems do. They generate citations by predicting what looks correct, not by confirming that a real document or database entry exists behind it. That’s why apparent precision can be misleading. A well-formatted citation doesn’t guarantee a verifiable source. When reliability depends on traceability, citations require more than appearance. They require #validation. ▶️ Why this matters for #mobility and #tolling In tolling, CITS, and mobility infrastructure, decisions depend on validated data. Architecture reviews often show that errors rarely arise from the AI model itself. Errors usually come from missing guardrails around it. A reference to an outdated rate, a missing policy, or a synthetic rule can cascade into compliance or billing errors. The same applies to citations: without post-generation checks, it is impossible to guarantee that what looks correct is, actually, correct. Governed #AI systems are beginning to apply multi-layer verification to ensure citation reliability: 🔸 Automated retrieval validation to cross-check that cited documents actually exist and match content; 🔸 RAG-based citation pipelines that link every answer to a validated, timestamped source; 🔸 Audit models (“AI judges”) that may use secondary LLMs or rule-based systems to verify the output of the first; 🔸 Human-in-the-loop review to resolve and "stamp" cases where factual or legal accuracy must be confirmed before publication or operational use. In regulated #infrastructure, citations are equivalent to proof of origin. Tolling systems must show that every tariff, policy, or enforcement action comes from an official source. Here, the AI system must prove that each claim or output can be traced to verified data. And, yes, each of these steps adds cost and complexity, but it also adds credibility. Reliability comes from validation logic, not appearance. ⚡ Leadership takeaway A trustworthy system is one that can prove its references. In regulated environments, this chain of proof is as important as the output or result itself. Transparency and verification are not optional. They are the foundation for trust, compliance, and resilience. Across sectors - from #mobility to #finance - the question is not whether AI can produce references. It certainly can. They questions is whether we can rely on them. 👉 Next Wednesday: Can Al Use All My Documents at Once? #AI #Leadership #Governance #Trust #Compliance #Audit

🚀 #𝗗𝗮𝘁𝗮𝗦𝗰𝗶𝗲𝗻𝗰𝗲𝗨𝗻𝗹𝗼𝗰𝗸𝗲𝗱: 𝗩𝗲𝗿𝗶𝗳𝗶𝗲𝗱 𝗔𝗜 𝗣𝗶𝗽𝗲𝗹𝗶𝗻𝗲𝘀 – 𝗕𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗧𝗿𝘂𝘀𝘁 𝘄𝗶𝘁𝗵 𝗠𝗼𝗱𝗲𝗹 𝗩𝗮𝗹𝗶𝗱𝗮𝘁𝗶𝗼𝗻 𝗟𝗼𝗼𝗽𝘀 Over the past few months, through my #DataScienceUnlocked series, I’ve been exploring how AI is reshaping industries — from RAG to RAAG, from Agentic AI to Multimodal systems. Each post dives into how data science is evolving beyond algorithms — towards intelligent, trustworthy systems that can reason, validate, and act. This week, let’s talk about something every AI practitioner eventually faces: 👉 How do we verify what AI produces? 🔍 What’s a Verified AI Pipeline? A verified pipeline doesn’t stop at generation — it goes one step further to validate every output before it’s accepted. It’s like having a quality control layer built right into your AI system. Instead of “Prompt → Output,” the new loop is: Generate → Validate → Feedback → Improve This shift ensures every AI-produced insight is accurate, consistent, and explainable — essential when decisions depend on them. ⚙️ 𝗛𝗼𝘄 𝗜𝘁 𝗪𝗼𝗿𝗸𝘀 (𝗪𝗶𝘁𝗵𝗼𝘂𝘁 𝘁𝗵𝗲 𝗕𝘂𝘇𝘇𝘄𝗼𝗿𝗱𝘀) Think of examples across domains: 🏦 Finance: AI-generated investment summaries verify that risk scores match the actual portfolio data. 🧑💻 Software: Code assistants auto-run unit tests or linters before suggesting solutions. 📰 Content Generation: News summarizers cross-check dates, facts, and names with reference data. 🧬 Healthcare: Automated summaries ensure percentages and patient counts align with original datasets. These validations use: Semantic similarity checks: Ensuring generated text matches the source meaning. Factual scoring: Smaller “checker” models judge whether outputs are grounded. Rule-based QA: Numeric or logical validations, like schema or threshold checks. Cross-model review: One model generates, another critiques. Together, they create audited intelligence — AI that double-checks itself. 🧠 𝗪𝗵𝘆 𝗜𝘁 𝗠𝗮𝘁𝘁𝗲𝗿𝘀 Large language models are probabilistic — meaning they sound confident, even when wrong. In critical workflows, that’s not acceptable. Verified pipelines bring determinism and traceability: Trust: You can trace every AI statement back to its source. Efficiency: Automated checks reduce manual reviews. Compliance: Outputs align with data, rules, and regulations. 🌐 𝗧𝗵𝗲 𝗕𝗶𝗴𝗴𝗲𝗿 𝗣𝗶𝗰𝘁𝘂𝗿𝗲 The next leap in AI isn’t just making models smarter — it’s making them accountable. Soon, every generative model will have a built-in validator — a silent reviewer ensuring that AI remains reliable, transparent, and responsible. #AI #DataScience #LLM #ResponsibleAI #TrustworthyAI #FutureofWork #DigitalTransformation #Innovation #Technology #DataScienceUnlocked

Some systems learn once and then repeat what they know. Others can look up new information as they go. The most advanced kind can plan, act, and improve after every step. The real progress happens when we combine all three so AI can understand, decide, and get things done in real time. Here's a quick breakdown of the three systems and how they differ: 1/ Model centric AI This is the traditional form of artificial intelligence. The model is trained on a fixed set of data and relies entirely on what it has already learned. It begins by defining the problem, collecting and labelling data, training and fine tuning the model, then deploying and evaluating its performance. When new information changes the patterns, it must be retrained. This works well for stable and predictable tasks such as image recognition, fraud detection, and sentiment analysis, but it cannot learn or adapt beyond the data it was trained on. 2/ Data centric AI (RAG systems) Retrieval Augmented Generation, often called RAG, changes the focus from what the model knows to what it can find. Facts are stored in an external database (or searched live from the internet) and retrieved dynamically whenever they are relevant. The model then uses that context to generate accurate, current answers and refines its performance through feedback. These systems do not rely on memory but look up information as needed. They're ideal for copilots, knowledge assistants, and enterprise search tools. While they are reactive and only intelligent when prompted, they are far more flexible than static models. 3/Agentic AI This emerging form of AI goes beyond responding to prompts. It can plan, take action, and learn from outcomes within defined boundaries. It starts with a goal, uses large language models for reasoning, connects to tools and data sources, and improves through reflection and feedback. These systems can execute workflows, coordinate with other agents, and adapt their approach based on results. Frameworks such as LangGraph, CrewAI, and MCP are already showing how this form of intelligence can work in practice. So where is all this taking us? Well, modern AI is increasingly combining these layers. Agentic systems use RAG for live data retrieval and governance layers for accountability and compliance. They can gather information, reason about it, take meaningful action, and record every step. This marks a shift from AI that predicts to AI that understands, decides, and evolves. So, to sum up, if traditional AI predicts from past data and RAG systems ground answers in live data, then Agentic AI acts and adapts in real time. And many are saying that the future belongs to AI that brings all three together to create predictive, grounded, and autonomous intelligence.

🔍 𝗧𝗵𝗲 𝗕𝗶𝘁𝘁𝗲𝗿 𝗧𝗿𝘂𝘁𝗵: 𝗔𝗜 𝗧𝗿𝗮𝗻𝘀𝗽𝗮𝗿𝗲𝗻𝗰𝘆 𝗜𝘀𝗻’𝘁 𝗪𝗵𝗮𝘁 𝗬𝗼𝘂 𝗧𝗵𝗶𝗻𝗸 𝗜𝘁 𝗜𝘀 Everyone’s talking about 𝗔𝗜 𝘁𝗿𝗮𝗻𝘀𝗽𝗮𝗿𝗲𝗻𝗰𝘆 but few truly understand what it means. The bitter truth? Most AI systems today are 𝘁𝗿𝗮𝗻𝘀𝗽𝗮𝗿𝗲𝗻𝘁 𝗼𝗻𝗹𝘆 𝗶𝗻 𝗣𝗼𝘄𝗲𝗿𝗣𝗼𝗶𝗻𝘁 𝘀𝗹𝗶𝗱𝗲𝘀, not in practice. Behind the buzzwords and polished dashboards, many AI models remain black boxes, trained on data we can’t trace, using logic we can’t explain, and producing results we can’t fully justify. Let’s uncover a few uncomfortable facts 👇 ⚠️ 𝗧𝗵𝗲 𝗠𝘆𝘁𝗵𝘀 𝘃𝘀. 𝗥𝗲𝗮𝗹𝗶𝘁𝘆 𝗠𝘆𝘁𝗵 𝟭: “We’re transparent because we use open-source AI.” 𝗥𝗲𝗮𝗹𝗶𝘁𝘆: Open code ≠ open understanding. True transparency requires documentation, audit trails, and interpretability. 𝗠𝘆𝘁𝗵 𝟮: “We disclose our data sources.” 𝗥𝗲𝗮𝗹𝗶𝘁𝘆: Without data lineage, consent, and bias evaluation, disclosure means nothing. 𝗠𝘆𝘁𝗵 𝟯: “Our AI is explainable, it shows confidence scores.” 𝗥𝗲𝗮𝗹𝗶𝘁𝘆: A confidence score doesn’t explain why the AI made that choice. 𝗠𝘆𝘁𝗵 𝟰: “Transparency slows innovation.” 𝗥𝗲𝗮𝗹𝗶𝘁𝘆: It prevents irresponsible innovation. The cost of opacity is trust, once lost, it’s gone. 💡 𝗧𝗵𝗲 𝗥𝗲𝗮𝗹 𝗠𝗲𝗮𝗻𝗶𝗻𝗴 𝗼𝗳 𝗧𝗿𝗮𝗻𝘀𝗽𝗮𝗿𝗲𝗻𝗰𝘆 True AI transparency means:   • 𝗞𝗻𝗼𝘄𝗶𝗻𝗴 𝗵𝗼𝘄 data is collected, processed, and used.   • 𝗨𝗻𝗱𝗲𝗿𝘀𝘁𝗮𝗻𝗱𝗶𝗻𝗴 𝘄𝗵𝘆 the model made a decision.   • 𝗧𝗿𝗮𝗰𝗸𝗶𝗻𝗴 𝘄𝗵𝗼 is accountable when it fails.   • 𝗗𝗶𝘀𝗰𝗹𝗼𝘀𝗶𝗻𝗴 𝘄𝗵𝗮𝘁 risks are involved openly. Transparency isn’t about revealing trade secrets; It’s about building trust, fairness, and accountability into the AI lifecycle. 🌍 𝗪𝗵𝘆 𝗜𝘁 𝗠𝗮𝘁𝘁𝗲𝗿𝘀 In an era where algorithms decide 𝘄𝗵𝗼 𝗴𝗲𝘁𝘀 𝗹𝗼𝗮𝗻𝘀, 𝗷𝗼𝗯𝘀, 𝗼𝗿 𝗲𝘃𝗲𝗻 𝗳𝗿𝗲𝗲𝗱𝗼𝗺, opacity isn’t just a technical flaw, it’s an 𝗲𝘁𝗵𝗶𝗰𝗮𝗹 𝗰𝗿𝗶𝘀𝗶𝘀. AI shouldn’t just be powerful. It must be 𝘂𝗻𝗱𝗲𝗿𝘀𝘁𝗮𝗻𝗱𝗮𝗯𝗹𝗲, 𝗮𝘂𝗱𝗶𝘁𝗮𝗯𝗹𝗲, 𝗮𝗻𝗱 𝗮𝗻𝘀𝘄𝗲𝗿𝗮𝗯𝗹𝗲. 🚀 𝗧𝗵𝗲 𝗪𝗮𝘆 𝗙𝗼𝗿𝘄𝗮𝗿𝗱 Demand 𝗔𝗜 𝗻𝘂𝘁𝗿𝗶𝘁𝗶𝗼𝗻 𝗹𝗮𝗯𝗲𝗹𝘀 clarity on data, purpose, and limitations. Support 𝗔𝗜 𝗴𝗼𝘃𝗲𝗿𝗻𝗮𝗻𝗰𝗲 𝗳𝗿𝗮𝗺𝗲𝘄𝗼𝗿𝗸𝘀 like ISO/IEC 42001 and the EU AI Act. Build systems that are 𝘁𝗿𝗮𝗻𝘀𝗽𝗮𝗿𝗲𝗻𝘁 𝗯𝘆 𝗱𝗲𝘀𝗶𝗴𝗻, not transparent by excuse. 🧩 𝗙𝗶𝗻𝗮𝗹 𝗧𝗵𝗼𝘂𝗴𝗵𝘁 The next time someone says their AI is “transparent,” Ask them, transparent to whom, and for what purpose? Because 𝗿𝗲𝗮𝗹 𝘁𝗿𝗮𝗻𝘀𝗽𝗮𝗿𝗲𝗻𝗰𝘆 𝗶𝘀𝗻’𝘁 𝗮𝗯𝗼𝘂𝘁 𝘀𝗵𝗼𝘄𝗶𝗻𝗴 𝗲𝘃𝗲𝗿𝘆𝘁𝗵𝗶𝗻𝗴. It’s about showing what truly matters. #TheBitterTruth #AITransparency #AIGovernance #AITrust #EthicalAI #CyberSecurity #AIRegulation #DigitalTrust #RiskManagement #WAIG #ResponsibleAI #IndianTechSociety #RisingTechVoices2025 #TechInfluencers2025 #LinkedInTopVoices #IndianLinkedInTechInfluencers #UKLinkedInTechInfluencers #Governance WAIG Institute & Research Center | AIforAll Global

The alarming headline is clear: “45% of Generative AI Agent Responses Contain Errors.” As an #AIBuilder, this data isn't just news; it's a direct call to action. The rapid pace of innovation must be matched by an unwavering commitment to #Reliability. 1. Reliability Must Be Our Default We can't settle for "it works" or "it's interesting." We must strive for solutions that: Work well. Do the right thing. Can be explained and defended. If a system delivers misleading or dangerous answers, innovation isn't a benefit—it’s a risk. Our first priority is transforming that risk into #Trust. 2. Controlled Knowledge is the Competitive Edge The root of most errors? Outdated or bad data. We cannot rely on the "generic web." To combat this, the key is the #ControlledCorpus: We must control our knowledge pipeline. Define reliable sources. Ensure our RAG/embedding architecture is filtering out "garbage" and delivering "gold." This isn't just engineering; it's #KnowledgeGovernance. 3. The AI Builder's New Role: The Governance Architect Our job extends beyond prompts and infrastructure. Today, we are Governors: Govern Sources & Data: Rigorously validating the knowledge base. Govern Response Quality: Implementing metrics for truthfulness and context relevance. Educate Users: Making it clear that AI is a powerful tool, not an "infallible authority." 4. This Is Our Greatest Opportunity If the industry average is a 45% error rate, it creates a massive gap. Vertical, controlled, and well-integrated AI solutions are the future. We can be the builders who step up and say: "Yes, generative AI exists, but the solution I build is more reliable, runs on a verified corpus, and operates through a controlled process." This is the differentiator that will define market leaders in Italy and beyond. 🇮🇹 Positioning for the Italian Market The Italian business landscape values precision, compliance, and control—especially with proprietary data. Positioning our solutions on Reliability, Verified Data, and Governance addresses their core concerns directly, transforming AI from a futuristic experiment into a de-risked strategic investment. Let's lead the way from Generative Hype to Generative Trust. What specific sector (e.g., Legal, Finance, Manufacturing) should we focus on first to deploy this Strategic Positioning Framework? #AIGovernance #EthicalAI #DigitalTransformation #Innovation #AffidabilitàAI #ItalianMarket

Problematic use of terminologies and reporting of analysis in this article. For reports like these, full disclosure and transparency of nuances observed in the use of AI tools is essential especially when it claims that it would save potential equivalent of 75000 days of manual human labour input in a year! https://lnkd.in/eKnVsQtt

Your AI Returns Too Much Information. Or Not Enough. Today’s risk is a quiet but damaging failure in knowledge-graph retrieval: I am talking about using the same retrieval granularity for every question, no matter what the question actually requires. The Risk Knowledge graphs can return information at different levels: • Nodes: “What is aspirin?” • Triplets:  “What does aspirin treat?” • Paths:  “How are aspirin and heart health connected?” • Subgraphs:  “What’s the full treatment pathway?” If a system chooses one granularity and apply it everywhere, that seem convenient and easy, but what could happen? The result: Simple questions get buried in noise. Complex questions get incomplete fragments. How This Shows Up in Your Business: Scenario 1: Simple question, overloaded retrieval Customer asks: “What’s the return policy?” Your system pulls the entire policy subgraph: international exceptions, damaged-goods flows, gift returns, timelines, and more. What you needed: one node with the basic return window. Your LLM’s context is now flooded with irrelevant tokens. Scenario 2: Complex question, insufficient retrieval Customer asks: “Can I return a gift shipped internationally if it arrived damaged?” Your system returns only the generic policy node. What you needed: a connected path across international shipping, damaged goods, gift returns, and exceptions. The AI answers incorrectly because it can’t see how these rules interact. Why It Matters: • Simple questions waste context budget: Entire org charts retrieved for “Who is the CEO?” • Complex ones break reasoning: Multi-hop logic needs connected context, not isolated triplets. • Answer quality drops: Too much context = noise; too little = missing logic. • Costs spike: Large subgraphs inflate token usage for no benefit. Mitigations: Use adaptive granularity • Factual lookups → Nodes • Relationship questions → Triplets • Connectivity questions → Paths • Multi-hop reasoning → Small subgraphs Compress before generation: • Turn triplets into natural language • Remove redundant properties • Summarize long chains Add governance: • Define retrieval rules per use case: support, recommendations, compliance, etc. • Track token usage, answer quality, and context utilization • A/B test fixed vs. adaptive strategies Bottom Line Not every question needs the same context. Retrieving too much leads to noise. Retrieving too little breaks reasoning. Match your retrieval granularity to the question, and your AI becomes more accurate, cheaper, and dramatically more reliable. #KnowledgeGraphs #GraphRAG #AIEngineering #LLMOptimization #InformationRetrieval #TokenEfficiency  

Topic of the data: Manually Labelled Data Here are a few articles by Karyna Naminas and Andy Muns to look through to give a broader overview: https://lnkd.in/gMd5_Wsd https://lnkd.in/g2ebtdDc  https://lnkd.in/g-cSM5xD Why manually-labelled data still rules in AI deployments: In today’s age of lightning-fast large language models and generative AI, it’s tempting to treat labelled data as a solved problem. But here’s the truth: human-driven, high-quality manual annotation remains foundational for trustworthy, scalable AI. Key reasons: -Labelled data serves as the ground truth that enables models to learn real world patterns. -In domain-specific environments (e.g., retail catalogs, supply-chain edge cases) nuance, context and business rules matter, only humans can reliably capture many of these subtleties. -Even with automation, hybrid workflows (human-in-the‐loop) are critical for quality, bias mitigation, and model alignment with business / ethical objectives. -Poor labelling = poor outcomes, aka garbage in = garbage out. A model trained on weak-labels or inconsistent annotation often fails to generalise, misses edge cases, or introduces hidden bias. Why this matters for me (and you): As someone leading AI strategy in high-stakes environments (e-commerce / retail / supply-chain), I live in the “last mile of model deployment” world.  Success hinges less on the flashy model architecture and more on the fidelity of the data that feeds it. Without a robust annotation layer, scaling becomes brittle, results become unpredictable, and business adoption stalls. Call to action: If you’re working on an AI initiative and haven’t built (or audited) your manual-labelling strategy yet, pause. Ask: Do we have clear annotation guidelines tied to business outcomes? Are edge-cases and domain nuance being surfaced and incorporated? Is there a feedback loop from production failures back into the labelled dataset? Are we measuring label quality (inter-annotator agreement, drift, bias) and improving it over time? Let’s build AI that isn’t just powerful, but also reliable, aligned and ready for the real world.

𝗪𝗵𝗲𝗻 𝗰𝗮𝗻 𝗱𝗮𝘁𝗮 𝘁𝗿𝘂𝗹𝘆 𝗯𝗲 𝗰𝗮𝗹𝗹𝗲𝗱 “𝗔𝗜-𝗿𝗲𝗮𝗱𝘆”? Readiness doesn’t mean perfection – it means predictability. Data needs to be well-documented, repeatable, and reusable. It’s like grammar in a language – it doesn’t guarantee beautiful sentences, but makes understanding possible. The same goes for data: we can only talk about true AI readiness when we know how it’s structured and connected. “Once key variables are identified, the next step is making data usable and reusable. Frameworks like AIDRIN and ODI help transform subpar data into structured, usable assets that accelerate AI development.” 📖 𝙍𝙚𝙖𝙙 𝙩𝙝𝙚 𝙛𝙪𝙡𝙡 𝙖𝙧𝙩𝙞𝙘𝙡𝙚 𝙤𝙣 𝙇𝙞𝙣𝙠𝙚𝙙𝙄𝙣: https://lnkd.in/dMtaN8Uq 🌐  𝙀𝙭𝙥𝙡𝙤𝙧𝙚 𝙥𝙧𝙚𝙫𝙞𝙤𝙪𝙨 𝙚𝙙𝙞𝙩𝙞𝙤𝙣𝙨 𝙤𝙣 𝙤𝙪𝙧 𝙬𝙚𝙗𝙨𝙞𝙩𝙚: https://lnkd.in/dTG-xVa2 #AIReadiness #DataGovernance #DataQuality #AIImplementation #DigitalTransformation #Evergo #BetterBusiness