Ashish Arora

𝗡𝗲𝘄 𝗩𝗶𝗱𝗲𝗼: 𝗖𝗼𝗺𝗽𝗼𝘀𝗶𝘁𝗲 𝗦𝗲𝗺𝗮𝗻𝘁𝗶𝗰 𝗠𝗼𝗱𝗲𝗹𝘀 𝗘𝘅𝗽𝗹𝗮𝗶𝗻𝗲𝗱 | 𝗗𝗶𝗿𝗲𝗰𝘁 𝗟𝗮𝗸𝗲 + 𝗜𝗺𝗽𝗼𝗿𝘁 𝗧𝗮𝗯𝗹𝗲𝘀 𝗶𝗻 𝗣𝗼𝘄𝗲𝗿 𝗕𝗜 (𝟮𝟬𝟮𝟲 𝗨𝗽𝗱𝗮𝘁𝗲): https://lnkd.in/gqUXEQwp One suggestion I saw on the channel was to include questions like 𝘄𝗵𝗮𝘁, 𝘄𝗵𝘆, 𝘄𝗵𝗲𝗿𝗲, 𝗵𝗼𝘄, 𝘄𝗵𝗼, etc. I’ve been trying that in my last two videos. Hope you like this approach. In this video, we will cover:  What is the Composite Semantic Model? Why use a Composite Semantic Model? Where are the options? How do you create a Composite Semantic Model?

38

Post 3: AI Governance vs. Data Governance — same team, different jobs If post 2 was the “why,” today is the “how these fit together.” A quick cheat sheet: 🧱 Data Governance = the foundation Policies, quality, lineage, privacy, retention, access. It makes sure the ingredients (data) are trustworthy, protected, and well-managed. 🧭 AI Governance = the guardrails for decisions made with models Use-case risk, model lifecycle controls, transparency, human oversight, monitoring. It ensures the recipe (models + usage) delivers outcomes we can stand behind. What’s different (in practice): Object of control: data assets vs. model + use case Failure modes: bad/unsafe data vs. harmful/biased/opaque model behavior Time horizon: periodic stewardship tasks vs. continuous monitoring (drift, incidents, policy changes) Accountability: data owners/stewards vs. product owners + model risk/AI governance Where they lock arms (the “handoffs”): 🤝 Intake: DG validates sourcing/consent; AIG validates purpose, risk tier, and human-in-the-loop. 🤝 Build: DG enforces access + minimization; AIG enforces documentation (model cards), testing, bias/robustness gates. 🤝 Deploy: DG ensures compliant data flows; AIG sets approval + escalation paths and defines KPIs/SLOs for behavior. 🤝 Operate: DG monitors data quality changes; AIG monitors drift, incidents, and retrain triggers with auditability. 🤝 Change: DG governs new data/retention updates; AIG re-assesses risk when scope, model, or context shifts. Bottom line: Data Governance keeps the inputs worthy of trust. AI Governance keeps the outcomes worthy of trust. You need both to scale AI responsibly without slowing innovation. How are these functions organized in your world—under one umbrella or separate teams that partner? What’s worked (or not)? 👇 #AIGovernance #DataGovernance #ResponsibleAI #ModelRisk #MLOps #AILeadership

343

9 ความคิดเห็น

We need to rethink data integrations. The traditional “set it and forget it” model of moving large volumes of data on fixed schedules is increasingly ill suited in an AI-driven world. Not because those patterns are wrong, but because they were designed for a very different set of outcomes. Data volumes are exploding. Unstructured data now dominates. And yet we’re still spending enormous effort moving data that adds little marginal value, while ignoring signals that could materially change how we govern, trust, and use data. Let’s be honest: many organizations are still running ETL jobs to support dashboards that haven’t been viewed in years. At the same time, some of the most valuable insights, buried in unstructured content, remain completely untouched. What’s missing isn’t integration tooling. It’s intent. Our integration processes need to become more selective, more contextual, and far more purpose-driven. In other words, they need to become more intelligent. This is where agents become very interesting. Not as a replacement for deterministic, event-driven pipelines - we already do plenty of that - but as a layer of intelligence that can reason about when integration actually matters. Instead of blindly moving data from source to destination, agentic integrations could decide: ✅ What is worth moving, ✅ When it’s worth moving, and ✅ Whether it even needs to move at all. Traditional data quality and policy checks can still apply in-flight. What I'm talking about here goes well beyond any event or trigger based integrations we have today. Agents can operate asynchronously to apply additional context, especially in situations where business intent or governance relevance isn’t obvious upfront. This is particularly relevant for unstructured data. We don’t need to dump all unstructured content into lakes or warehouses. What we need is the metadata that matters, selectively extracted, purposefully integrated, and evaluated through a governance lens. At scale, agents are one of the few viable ways to make that determination dynamically. The shift isn’t from ETL to agents - or at least, not entirely. It’s from volume-driven integration to value-driven integration. And that distinction matters more than we may be ready to admit. Maybe some are already working on this? Thoughts? #dataintegration #etl #datagovernance

36

8 ความคิดเห็น

Here's what no RAG tutorial tells you Post #4 of 4 The retrieval method you choose is maybe 30% of the problem. The pipeline you build around it is the other 70%. After designing RAG systems across documents, code, financial data, and research corpora — here are the 6 pipeline patterns that actually work in production: ━━━━━━━━━━━━━━━━━━ ✅ 1. Simple Document Q&A (150–400ms) Query → Dense vector search (top-50) → Re-rank (top-5) → LLM with citations Covers 80% of FAQ, policy, and document Q&A. Start here. Add hybrid only when keyword recall is low. 🏢 2. Enterprise Search — SharePoint, intranets (200–500ms) Query → Permission filter (metadata pre-filter) → Hybrid search → RRF merge → Parent-doc expansion → Re-rank → LLM Critical: permissions filter BEFORE retrieval, not after. Post-filtering at scale is both slow and a security risk. 🔬 3. Research Deep-Dive (500ms–1.5s) Query → HyDE (generate hypothetical answer) → Dense search on abstracts → Parent expansion → Citation graph traversal → Re-rank → LLM synthesis Latency is worth it. This is research-quality retrieval. The citation graph surfaces related work the query alone would never find. 📊 4. Financial / Structured Data (300–700ms) Query → Intent classifier: prose or numeric? → Prose: hybrid + re-rank → Numeric: Text2SQL + execute → If both: merge → LLM synthesises The routing step is everything. Misrouting a revenue question to vector search produces hallucinated numbers. The classifier is not optional. 💻 5. Code Intelligence — IDE and dev tools (100–300ms) Query → Sparse search (exact identifiers) + Dense search (semantic) → RRF merge → Call graph traversal (depth 1–2) → Parent-doc (fetch full class) → LLM with file + line citations Always cite file path and line number. A code answer without a source location is not actionable. 💬 6. Conversational / Multi-turn RAG (400–800ms) New message → Query rewriting (condense history into standalone query) → Hybrid search → Re-rank → LLM with context + history → Update memory Query rewriting is the step 90% of teams skip. Pronouns and implicit references from previous turns destroy retrieval quality. The rewrite replaces "it" and "that" before any search happens. ━━━━━━━━━━━━━━━━━━ This wraps up my 4-part series on RAG architecture for real-world deployments. If you found this useful — the best thing you can do is share it with someone building an AI product right now. Most teams are still treating RAG as a single generic pattern. Happy to go deeper on any of these pipelines. Drop your use case in the comments. #RAG #AIArchitecture #EnterpriseAI #GenerativeAI #LLM #ProductAI #AIEngineering

14

The real test of AI governance is not deployment. It is capital allocation. When AI influences pricing, supply allocation, risk exposure, or investment timing, boards are no longer governing systems. They are governing consequences. The question executives should ask is simple: If the AI system is wrong tomorrow, who absorbs the economic outcome? If that answer isn’t clear, governance hasn’t caught up with capability. Architecture determines durability. Authority determines trust. #BoardOversight #EnterpriseAI #CapitalAllocation

3

𝐖𝐡𝐚𝐭 𝐢𝐟 11.11 𝐚𝐧𝐝 12.12 𝐚𝐫𝐞 𝐠𝐫𝐨𝐰𝐢𝐧𝐠 𝐭𝐡𝐞 𝐦𝐚𝐫𝐤𝐞𝐭, 𝐛𝐮𝐭 𝐧𝐨𝐭 𝐲𝐨𝐮𝐫 𝐎𝐟𝐟𝐢𝐜𝐢𝐚𝐥 𝐒𝐭𝐨𝐫𝐞? This latest dataset from Magpie IQ on 𝐈𝐧𝐝𝐨𝐧𝐞𝐬𝐢𝐚’𝐬 𝐜𝐨𝐧𝐝𝐢𝐦𝐞𝐧𝐭𝐬 category in H2 2025 shows why. In 𝐃𝐞𝐜𝐞𝐦𝐛𝐞𝐫, total category GMV soared to Rp90.7bn. But 𝐎𝐟𝐟𝐢𝐜𝐢𝐚𝐥 𝐒𝐭𝐨𝐫𝐞𝐬 captured only 39.4%, while 𝐦𝐚𝐫𝐤𝐞𝐭𝐩𝐥𝐚𝐜𝐞 𝐬𝐞𝐥𝐥𝐞𝐫𝐬 took 60.6%. It reveals a 𝐛𝐫𝐮𝐭𝐚𝐥 𝐭𝐫𝐮𝐭𝐡: the category’s biggest month came with a collapse in Official Store share. The missing link is C2C. Not because it is invisible in the market, but because it is still invisible to many brands. And that blind spot can cost brands millions in diverted revenue and 𝐜𝐮𝐬𝐭𝐨𝐦𝐞𝐫 𝐥𝐨𝐲𝐚𝐥𝐭𝐲. This is not just a revenue issue. It is a market control issue. When category demand peaks, brands should be in their strongest position. But if more of that upside flows through marketplace sellers, 𝐭𝐡𝐫𝐞𝐞 𝐪𝐮𝐞𝐬𝐭𝐢𝐨𝐧𝐬 𝐦𝐚𝐭𝐭𝐞𝐫: * Which C2C sellers captured the demand that Official Stores expected to win?   * What allowed them to take that share in the first place?   * Did they simply outperform Official Stores on visibility and conversion? When your category peaks, you cannot afford to be blind. 𝐓𝐡𝐚𝐭 𝐢𝐬 𝐰𝐡𝐞𝐫𝐞 𝐌𝐚𝐠𝐩𝐢𝐞 𝐈𝐐 𝐜𝐨𝐦𝐞𝐬 𝐢𝐧. We do not just show the numbers. We help brands see where capture shifts, who is driving it, and what needs to change before the next peak. The peak season should be your peak season. Comment below or 𝐃𝐌 𝐦𝐞 to learn how Magpie IQ can give you total market visibility before the next big sale.

10

1 ความคิดเห็น

Some more fun with TypeSafe AI's new Jev model! This time, I used it to evaluate the quality of data sources and claims in AI-led deep research. First, some context. A few weeks back, I wrote about common AI failure modes (e.g. hallucination, context rot, instruction drift), something Gary Ang, PhD and I have been exploring in recent months. In that post, I shared recent examples of big firms like EY, KPMG and Deloitte having to retract reports that contained hallucinated sources, citations and facts. With more people using agents to automate market research, checking the quality, authority, and relevance of data sources matters more than ever. Human-led checks are ideal, but they take time. So the trick is finding the right balance between human and AI-led evals. So where does Jev fit in? I ran a quick experiment using Jev to evaluate the sources and claims in an AI-generated research report. 🧪 THE EXPERIMENT The input was a 5,100-word Gemini Deep Research report about the global L&D market. The report made 177 cited claims drawing on 57 unique sources. To prepare the data for Jev, a script parsed the report extracting each claim and its citation, then fetching the cited web page, locating the relevant stats and quotes, and pulling supporting signals like domain age and traffic rank. Jev then answered the following questions, each with a probability score: 1. What kind of publisher is this (content farm, vendor, trade press, research org, etc)? 2. Is this likely SEO (i.e. content marketing) or AI-generated content? 3. Does the publisher have a commercial interest in this market? 4. For each number or stat in the report, was it traceable to the source, and if so, was the context consistent between the Gemini report and source. 📊 THE RESULTS → Jev processed 5,188,626 input tokens, finished in about 70 seconds, and cost me just $0.21 in inference. → Based on Jev's judgments, 95% of the 177 cited claims pointed to sources below trade-press authority (i.e. not very trustworthy). This was the major takeaway. I expected to find more hallucinated, misattributed or misinterpreted facts from the Gemini report (and there were some). But outright fabrication was rarer than I expected. The bigger problem was where the claims came from: content farms, vendor marketing and paid or promotional content. → Jev also caught a handful of stats that didn't appear anywhere on the cited page. One was a headline estimate of a sub-sector's TAM (a pretty important number), and none of its figures were on the page it cited (founders, beware of using deep research for your pitch deck!) Overall, I wouldn't say Jev is smarter than other leading LLM's like Claude. Where it shines is speed and token efficiency: it completed this task in a fraction of the time and cost. Also, reading the output as probabilities, rather than prose you need to read and parse feels surprisingly intuitive for this kind of task.

8

1 ความคิดเห็น

Real transformation happens when we shift from building dashboards to solving real problems for real roles. From requests → to outcomes. From data delivery → to decision support. The future of data teams isn’t being a report factory. It’s becoming partners who deeply understand business personas, workflows, and moments that matter — and design solutions around them. This is the evolution data teams need: Less “dashboard concierge,” more “people-first problem solving.”

5

3 ความคิดเห็น