Krishanu Agrawal

Jaya Gupta on the importance of decision context and why incumbent systems of records must retool to capture this decision data or face losing out to new products that own the orchestration path. — Warehouses do have a time-based view. You can query historical snapshots, track how metrics changed, and compare state across periods. But warehouses receive data via ETL after decisions are made. By the time data lands in Snowflake, the decision context is gone. A system that only sees reads, after the fact, can’t be the system of record for decision lineage. It can tell you what happened, but it can’t tell you why. … Systems of agents startups have a structural advantage: they’re in the orchestration path. When an agent triages an escalation, responds to an incident, or decides on a discount, it pulls context from multiple systems, evaluates rules, resolves conflicts, and acts. The orchestration layer sees the full picture: what inputs were gathered, what policies applied, what exceptions were granted, and why. Because it’s executing the workflow, it can capture that context at decision time – not after the fact via ETL, but in the moment, as a first-class record.

4

🚀 AI growth is accelerating. 💧 So is the conversation about water, energy, and community impact. The future of AI won’t be determined solely by model performance. It will also be determined by whether communities trust the infrastructure that powers it. That’s why I’m passionately building xOmni.io. Not to argue for or against AI. To help measure what matters: 💧 Water ⚡ Energy 🌎 Environmental impact 🏘️ Community outcomes Trust isn’t built through promises. It’s built through transparency. xOmni #AI #DataCenters #CommunityTrust #Sustainability #WaterStewardship #xOmni

2

After months of exploring AI implementation at Assignar, I’ve realised something crucial that most businesses miss: having access to AI tools like ChatGPT or Claude isn’t enough. Here’s the reality - while these tools are fantastic for one-off questions, they become truly transformative only when they understand YOUR business. Your most valuable employee isn’t the one who just started yesterday. It’s the person who knows your customers, understands your processes, and remembers that support ticket from six months ago that solved a similar problem. The same applies to AI. For AI to deliver real value in an enterprise setting, it needs access to: • Your support ticket history • Product usage patterns • Internal documentation • Customer interactions This is where things get technical (but stay with me). All this knowledge needs to be organized in what’s called a RAG database - essentially a smart filing system that AI can search through. Then you need a way for the AI to actually access this information. Here’s what surprised me: Currently, Claude.ai is the only mainstream chat interface that has built-in MCP (Model Context Protocol) - think of it as a universal adapter that lets AI tools connect to your company’s data securely. This means that if you use Claude.ai as your corporate LLM interface tool, you don’t need to build and maintain your own. The bottom line for business leaders: If you’re serious about AI transformation, don’t just think about which AI tool to use. Think about how to give that tool access to your institutional knowledge. The companies that figure this out will have a massive competitive advantage. We’re still early in this journey at Assignar, but the potential is clear. AI without context is just a smart calculator. AI with your company’s full context? That’s a game-changer.

23

1 Comment

One question frequently arises in enterprise AI discussions: “How do we use GPT-like intelligence without compromising data security?” For many organizations, this is not merely a tooling question; it revolves around design, governance, and trust. Public LLMs have demonstrated the potential of this technology. The real challenge lies in translating that capability within the confines of security controls, compliance boundaries, and existing data ecosystems. This often leads to discussions about offline versus online GPT approaches, which is rarely a binary choice. At a high level: - Offline/private GPTs offer stronger control over sensitive data, access, and auditability. - Online GPTs enable faster innovation cycles and broader language capabilities. What is most important is ensuring fit for purpose—considering data sensitivity, regulatory expectations, and the types of decisions being supported. A well-designed in-house GPT functions as a secure intelligence layer rather than a standalone chatbot: - It operates within the organization’s controlled environment. - Access is role-based and governed. - Responses are based on approved enterprise data. - Outputs support decisions without bypassing processes. When constructed this way, internal GPTs do not replace human judgment. Instead, they reduce friction, enhance consistency, and empower leaders and teams to act swiftly and confidently. At Canara HSBC we have been closely involved in shaping this space and translating these principles into a practical, enterprise-grade solution. We are nearing the point of bringing this to fruition and will share more updates as we progress. Happy to habe as many inputs.

27

New spec to bridge semantic definitions between AI and BI. And a great list of partners in the mix.

1

🚀 Just finished reading “Graph of Thoughts: Solving Elaborate Problems with Large Language Models” (AAAI-24) — and it really made me think. LLMs are learning to reason — not just respond. Here’s how reasoning has evolved, from stateless outputs to fully stateful, graph-structured thought 👇 ⚙️ Basic Input–Output — Stateless Why: The model processes the prompt once and produces an output — no internal reasoning steps are remembered. Example: Q: “Plan a trip to Japan.” A: “Visit Tokyo and Kyoto.” A direct answer — quick, but shallow. 🔗 Chain of Thought (CoT) — Minimally Stateful Why: Keeps a linear state — one reasoning path, step by step, but can’t explore alternatives or backtrack. Example: “Fly to Tokyo → Stay 3 days → Visit Kyoto → Return.” A clear sequence, but no flexibility — one reasoning chain. 🌲 Tree of Thoughts (ToT) — Partially Stateful Why: Maintains multiple reasoning branches (different possible trip plans) and picks the best after exploring options. Example: Path A: Tokyo + Kyoto (culture-focused) Path B: Tokyo + Okinawa (relaxation-focused) It evaluates both and chooses one — partially stateful, but branches don’t interact. 🌐 Graph of Thoughts (GoT) — Fully Stateful Why: Builds a network of connected ideas where different reasoning paths interact, merge, and update each other. Example: While planning a trip, the model links multiple factors: 🏙️ Destinations (Tokyo, Kyoto, Okinawa) 💰 Budget (flights ↔ hotels ↔ attractions) 🌦️ Weather (affects destination choices) 📅 Schedule (updates based on costs and weather) The model dynamically adjusts one idea based on another — e.g., “If it rains in Kyoto, shift to Okinawa; update budget and timing accordingly.” That’s graph reasoning — interconnected states influencing each other toward an optimal, consistent plan. 🧠 What struck me most is how GoT mirrors human reasoning — refining ideas through feedback, context, and connection. From chains → trees → graphs, LLMs are evolving from linear logic to adaptive, interconnected thought. #AAAI24 #GraphOfThoughts #LLM #AIResearch #Reasoning #MachineLearning #ArtificialIntelligence Reference: https://lnkd.in/gCkqjvPi

14

"GenAI interpretability gets less attention than the constant deluge of model releases, but it is arguably more important." I completely agree with this point from the recent Urgency of Interpretability blog by Dario Amodei. As foundation models evolve rapidly, the lack of interpretability can become a major barrier to safe, compliant, and large-scale adoption—especially in finance. To build trust, we must be able to explain model outputs clearly and reliably. I’m excited to speak on this topic at the upcoming AI4 2025 conference, where I’ll present our latest work: “Beyond the Black Box: LLM Interpretability in Finance.” I’ll be sharing how we’re getting into the "brain" of LLMs to uncover meaningful financial features and apply them across key use cases like trading and sentiment analysis. Thankful to my co-author, Ariye Shater, for his collaboration on this work. Links: - Paper Link: https://lnkd.in/exrmXh97 - Conference: https://ai4.io/finance/ - Dario's Blog: https://lnkd.in/ePDp-x-W #GenAI #Interpretability #Finance #LLMs #AI4Conference #SparseAutoencoders #Anthropic #ResponsibleAI #MechanisticInterpretability

87

I feel like we are seeing a critical market correction in AI. Most of us are now all too familiar with the recent MIT study, which found that 95% of enterprise generative AI projects are failing to achieve their objectives (https://lnkd.in/eavspBQR ). For me, this isn't a deterrent—it's a directive. The era of blind experimentation is over. The path to success, especially in a trust-centric industry like insurance, isn't about generalized models. It's about a strategic pivot to pragmatic, ROI-driven solutions that supercharge our workforce and agents. PwC's data shows AI-exposed industries have nearly three times higher growth in revenue per employee. This proves the real value lies in amplifying human potential. To be a leader driving genuine transformation, you must prioritize commitment to the workforce and responsibility, because trust is our currency.  I think we are ready for this future, and those of us who have taken our time to really consider the technology as an amplifier of what we already have (human capital), will understand the path to material ROI is through our human IP, not around it. https://lnkd.in/ecvAM8bq

59

9 Comments