Navigating Complex Systems

A big question in software is what happens to the systems of record in a world of AI Agents. Do they go away? Do they just become databases? Or do they become more powerful? I’d argue that they’re just as powerful as ever, if not more powerful, in a world of 100X more interactions with software. The purpose of your system of record (whether it’s ERP, CRM, ITSM, or a document management system) is to hold the data and manage the workflows around the most important areas of your business: your customer commitments, leads, revenue figures, inventory, IP, product research, supply chain, and more. Importantly, you want the data and workflows in these systems operate in deterministic ways. When you ask a question like “what is my revenue,” you need the precise answer. When you move a lead from one stage to another, you can’t afford for it to get dropped. When you update your inventory, you can’t have it change inadvertently. Getting the data, permissions, access controls, business logic, and workflows right, every single time, is critical. On the other hand, AI Agents operate in a world of non-deterministic actions. What makes them so powerful is they can adapt to entirely new instructions on the fly, use judgment to perform actions, and operate on troves of unstructured information and decisions. When you ask an AI Agent to research and summarize a set of documents, it will produce a slightly different answer every single time - and in most use-cases for AI Agents, this is a feature, not a bug. Just as you wouldn’t ask the world’s smartest human to memorize every piece of inventory you have, or all of the permissions of every information that employees should have access (with their specific access controls) to, you similarly won’t ask AI Agents to do that in the future. This is where the separation of duties comes into play. AI Agents will be doing non-deterministic actions (like generating a sales plan, responding to a customer, or writing code), and deterministic systems will be for remembering those actions and incorporating them across a variety of workflows. In fact, in a world of AI Agents running around doing autonomous tasks 24/7, in parallel, and at unlimited scale, the role that these systems of record play will likely be even more important. Getting this relationship down is going to be key to the future of the enterprise IT stack.

Before being CEO at Dimmo, I sold into large enterprise accounts (10k-300k employees). Here's how I consistently broke into accounts: Bare with me - this is a long, tactical post. 1. Created an Org Chat I wanted to understand the lay of the land from a people perspective. → C-Level → Any relevant VP → All Directors → Relevant Managers → Relevant Practitioners 2. Created an Account Plan. An account plan is key for large accounts and top prospects. → Overview → Investor Relations → 10k Information (if public) → Growth Insights → Recent news (are they doing well? bad?) → Job openings → Challenges → Concerns → Risk → How they make money → How do they go to market? → Their competitors → Their industry trends → Have ChatGPT summarize this info & why they're a good fit There's more you can do, but this is a great start. 3. Persona Information I did deep research on people - at a minimum Director level and above. → Podcast features → Article/Blog features → Accomplishments from current/previous roles → Video features → Twitter info → LinkedIn posts/likes → Any other relevant info when you type in their name and title in Google or ChatGPT *You can also use ChatGPT prompts to get this information and create a hypothesis as to why they would care 4. Use Multiple Channels Do a multi-channel approach: → LinkedIn → Email → Cold Calls → In-person events → Your current network Within these areas, get creative. 5. Signals I focused on 'warm' areas first. → Job changes → Industry changes → Intent (though most of this is eh) → People who open emails multiple times → People that click on links → Companies that are in growth mode Do this for your top 10 accounts & I promise you have a high chance of breaking into the account. Here are technologies that help understand who's in market: https://lnkd.in/e_J6witn Also, many of these technologies help me with account research: https://lnkd.in/dHmNdQt6

If you’re an AI engineer, product builder, or researcher- understanding how to specialize LLMs for domain-specific tasks is no longer optional. As foundation models grow more capable, the real differentiator will be: how well can you tailor them to your domain, use case, or user? Here’s a comprehensive breakdown of the 3-tiered landscape of Domain Specialization of LLMs. 1️⃣ External Augmentation (Black Box) No changes to the model weights, just enhancing what the model sees or does. → Domain Knowledge Augmentation Explicit: Feeding domain-rich documents (e.g. PDFs, policies, manuals) through RAG pipelines. Implicit: Allowing the LLM to infer domain norms from previous corpora without direct supervision. → Domain Tool Augmentation LLMs call tools: Use function calling or MCP to let LLMs fetch real-time domain data (e.g. stock prices, medical info). LLMs embodied in tools: Think of copilots embedded within design, coding, or analytics tools. Here, LLMs become a domain-native interface. 2️⃣ Prompt Crafting (Grey Box) We don’t change the model, but we engineer how we interact with it. → Discrete Prompting Zero-shot: The model generates without seeing examples. Few-shot: Handpicked examples are given inline. → Continuous Prompting Task-dependent: Prompts optimized per task (e.g. summarization vs. classification). Instance-dependent: Prompts tuned per input using techniques like Prefix-tuning or in-context gradient descent. 3️⃣ Model Fine-tuning (White Box) This is where the real domain injection happens, modifying weights. → Adapter-based Fine-tuning Neutral Adapters: Plug-in layers trained separately to inject new knowledge. Low-Rank Adapters (LoRA): Efficient parameter updates with minimal compute cost. Integrated Frameworks: Architectures that support multiple adapters across tasks and domains. → Task-oriented Fine-tuning Instruction-based: Datasets like FLAN or Self-Instruct used to tune the model for task following. Partial Knowledge Update: Selective weight updates focused on new domain knowledge without catastrophic forgetting. My two cents as someone building AI tools and advising enterprises: 🫰 Choosing the right specialization method isn’t just about performance, it’s about control, cost, and context. 🫰 If you’re in high-risk or regulated industries, white-box fine-tuning gives you interpretability and auditability. 🫰 If you’re shipping fast or dealing with changing data, black-box RAG and tool-augmentation might be more agile. 🫰 And if you’re stuck in between? Prompt engineering can give you 80% of the result with 20% of the effort. Save this for later if you’re designing domain-aware AI systems. Follow me (Aishwarya Srinivasan) for more AI insights!

What’s the fastest way to compress your sales cycle when selling large deals to Enterprise accounts? Here's what you DON'T do: - meet with department heads - demo your product and get them excited - have them attempt to sell upstream Instead, try using my Yo-yo sales framework. Here’s what it looks like. Step #1: Connect with power first, understand the most important problem(s) they need to solve, why they want to change, and the costs of not changing.If it’s a problem you can help solve, get their sponsorship for your discovery process. Ghost write a note they can send to their team to begin discovery Step #2: Meet with their department heads, get the real deal on what their day to day challenges look like, and grab some quotes which capture their biggest pain points. Seek to understand, not sell. Dig deep - understand the people, processes, and steps involved in doing the painful things which you can help improve or automate. Step #3: Package everything you’ve uncovered into an Executive Summary and Business Case and deliver back to Exec Sponsor and their team. Demo only what’s relevant. Close. Power compresses deals. This insight was key to me becoming the number one AE at Salesforce. It's how I went from averaging $240K/year my first four years to averaging $720K my final four years (including two years cracking 7 figures). Want to learn my process for closing huge enterprise deals? Tune-in to my full conversation on the 30 Minutes to President's Club podcast. Listen here: https://lnkd.in/gCJ4x7HZ

One thing I’ve learned building AI powered automation for clients: Deterministic and agentic automation do not ask for the same kind of patience. On paper, they both look like “a workflow”. In real life, they feel very different. Deterministic automation is the rules-based kind. If X happens, do Y. If a field is blank, stop. If a deal is in Stage 3, notify this person. You reach for this when: • The rules are clear • The data is structured • The same input should always produce the same output You spend most of your time defining the logic and testing a few scenarios. Once it works, it’s usually stable until something upstream changes. Agentic or AI powered automation is different. You’re not just routing data. You’re asking for judgment. Things like: • Summarize this email • Decide which team should handle this request • Draft a response that fits these guidelines • Classify this lead based on what they wrote Small input changes can shift the output. So the work changes too. You’re not just connecting steps. You’re shaping behavior. That means: • More rounds of testing with real examples • More time tightening prompts and instructions • More clarity on what “good” looks like and how to make it repeatable You do get to a stable version. It just takes more iteration to earn that stability. Neither type is “better”. They just shine in different places. • Clear rules, strict outcomes, predictable paths → deterministic • Messy inputs, natural language, prioritization, judgment → agentic The mistake is expecting agentic automation to behave like deterministic automation on day one. It can do more, but it also asks more from you: More patience. More examples. More care with inputs. Once you accept that, you start building with clearer expectations and a lot less frustration.

Still running 2025 problems on 1995 logic? They’ll say, “Software is predictable, just follow the spec.” They’ll argue, “Executives can’t bet on ‘maybe’.” That mindset made sense when code was linear and deterministic. 🔴 Traditional (Deterministic) Software 1. Rules in → Rules out Same input always yields the same output. 2. Fixed pathways Flowcharts end where they start, nothing new is learned. 3. Change occurs by release cycle Months of scoping, testing, and sign-offs before value appears. 4. Zero tolerance for ambiguity Anything less than 100 % certainty is marked as failure. 🟢 AI (Adaptive & Probabilistic) 1. Data in → Confidence scores out Outputs are likelihoods, not guarantees, letting you act on emerging signals. 2. Continuous learning Models refine themselves as new data streams in. 3. Real-time iteration Small pilots pivot daily, compounding gains while old projects wait for approval. 4. Risk managed by guardrails Governance shapes behaviour without strangling speed. 🚨 Where Linear Logic Fails Today • Yesterday’s KPIs hard-code blind spots. • Binary pass/fail gates kill novel insights. • Waterfall sign-offs stall momentum while markets shift. • Teams freeze when the model returns 73 % instead of 100 %. ✅ Why Probabilistic Leadership Wins • Acts on confidence intervals while rivals chase certainty. • Scales value through experimentation loops, not megaprojects. • Embeds safeguards at the data and model layer, freeing talent to innovate. • Turns “maybe” into competitive edge by learning faster than the environment changes. Paper maps give one static route. GPS plots optional paths, updates with live traffic, and recalculates when reality intrudes. Only one model survives congestion. This isn’t a faster version of the old waterfall, it renders waterfall obsolete. Prediction: Boards that demand linear absolutes will watch adaptive competitors erode their margins and poach their best people. Ready to steer your organisation by probabilities instead of certainties—or content to follow a map that no longer matches the terrain?

Interviewer: "Your AI chatbot receives 1 million requests every day. How would you reduce LLM costs without affecting answer quality?" Answer: The first thing I would not do is switch to a smaller model. Cost optimization should happen across the entire pipeline, not just at the model layer. 1. Introduce Semantic Caching Many users ask the same question in different ways. For example: - "What's your refund policy?" - "Can I get my money back?" - "How do refunds work?" Although the wording is different, the intent is the same. By storing embeddings of previous queries, we can retrieve a previously generated answer when a new query is semantically similar, avoiding another LLM call. Tools: Redis + vector search, FAISS, Qdrant, Pinecone. 2. Use Prompt Caching A large portion of prompts often contains static content such as: - System instructions - Company policies - Tool descriptions - RAG instructions Instead of sending these repeatedly, use provider-supported prompt caching where available. This reduces both input tokens and latency. 3. Multi-Model Routing Not every request requires the most expensive model. For example: - FAQs → Small LLM (Llama 3.1 8B, Gemma) - Summarization → Medium model - Complex reasoning or coding → GPT-5, Claude, or another high-end model A lightweight classifier or router can determine which model should handle each request. 4. Improve Retrieval Before Generation If you're using RAG, better retrieval means the LLM receives cleaner context. Focus on: - Better chunking - Hybrid search - Cross-encoder reranking - Query rewriting - Metadata filtering When the context is highly relevant, even smaller models can produce excellent answers. 5. Reduce Token Usage Every token costs money. Optimize by: - Compressing retrieved context - Removing duplicate chunks - Retrieving only the Top-K relevant documents - Limiting conversation history - Summarizing long chat histories instead of sending everything Reducing unnecessary tokens lowers both cost and response time. 6. Batch Non-Real-Time Requests Tasks such as document summarization, report generation, or data extraction don't always need immediate responses. Batching these requests improves throughput and reduces infrastructure costs. 7. Fine-Tune Small Models for Repetitive Tasks If a task is highly repetitive, such as intent classification, entity extraction, or support categorization, a fine-tuned smaller model can replace a large general-purpose LLM. This improves both speed and cost efficiency. 8. Continuously Monitor Cost and Quality Optimization is an ongoing process. Track metrics such as: - Cost per request - Token consumption - Cache hit rate - Latency - Model routing distribution - User satisfaction - Task success rate The goal is to reduce cost without degrading answer quality. Follow Sneha Vijaykumar for more...😊 #ai #llm #rag #aiengineer #interview #preparation #datascience

Exciting Innovation in AI Ranking Systems: CoRanking Approach! I'm thrilled to share a groundbreaking advancement in information retrieval technology that addresses a critical efficiency challenge in today's search systems. Researchers from Renmin University of China and Baidu have developed "CoRanking," a novel collaborative ranking framework that combines small and large language models to achieve both efficient and effective passage ranking. >> The Problem CoRanking Solves Large Language Models (LLMs) like GPT-4 have demonstrated superior listwise ranking performance for search results. However, this comes at a significant cost: - Massive computational overhead from large-scale parameters - Inefficient sliding window processes that can't be parallelized - High latency that limits real-world applications >> The Innovative Solution CoRanking introduces a three-stage collaborative framework: 1. Small Listwise Reranker (SLR): A 3B-parameter model pre-ranks all candidate passages, bringing relevant ones to the top-20 positions 2. Passage Order Adjuster (POA): Trained via reinforcement learning (specifically Direct Preference Optimization), this component reorders the top passages to align with the LLM's positional preferences 3. LLM Listwise Reranker (LLR): A 72B-parameter model that only needs to rerank the adjusted top-20 passages >> Technical Implementation Details The framework addresses a subtle but critical challenge: LLMs have significant positional biases when processing input passages. The POA component specifically tackles this through: - A significance-aware selection strategy (S³) that builds high-quality preference pairs - Human-label-enhanced ranking construction (HRC) that improves training quality - DPO optimization that aligns passage order with LLM preferences >> Impressive Results Extensive experiments across TREC DL, BEIR, and BRIGHT benchmarks demonstrate: - 70% reduction in ranking latency compared to pure LLM approaches - Superior effectiveness with higher NDCG@10 scores than using only LLM rerankers - Consistent performance across different LLM backends (tested with Qwen2.5-72B, GPT-4o, and DeepSeek-V3) This research represents a significant step forward in making advanced ranking systems practical for real-world applications. The collaborative approach between small and large models could become a blueprint for other AI systems seeking to balance efficiency and effectiveness.

Can STRATEGY learn anything from QUANTUM MECHANICS? Quantum mechanics offers valuable insights for strategic leadership in today's complex and uncertain business environment. Here's how we can apply quantum principles to enhance our leadership approach: 1]. EMBRACING UNCERTAINTY AND POSSIBILITY In quantum mechanics, particles exist in multiple states simultaneously until observed. Similarly, strategic leaders must embrace uncertainty and consider multiple possibilities. Instead of rigid, deterministic planning, we should: - Envision multiple potential outcomes for any situation - Explore diverse approaches with input from various stakeholders - Maintain flexibility to pivot as circumstances evolve This "superposition" mindset allows us to thrive on uncertainty and foster innovation at the "edge of chaos". 2]. THE POWER OF OBSERVATION AND INTENTION Just as observing quantum particles affects their state, a leader's focus shapes organizational reality. We must be mindful of our "observer effect" by: - Cultivating awareness of our perceptual biases - Intentionally creating a positive organizational culture - Balancing focus between efficiency (exploiting) and effectiveness (exploring) Our attention and expectations have ripple effects throughout the organization. 3]. INTERCONNECTEDNESS AND EMERGENCE Quantum entanglement demonstrates the interconnected nature of particles. In leadership, this translates to: - Fostering strong relationships and networks within teams - Recognizing that small actions can have far-reaching impacts - Allowing for bottom-up, self-organizing structures to emerge By cultivating a high "connectivity quotient," we can create teams that perform beyond the sum of their parts. 4]. ADAPTING TO COMPLEXITY Quantum uncertainty challenges traditional, linear planning. To lead effectively in complex systems: - Adopt an adaptive, learning-oriented approach to strategy - Encourage experimentation and "quantum tunneling" to overcome barriers - Focus on creating conditions for innovation rather than rigid objectives. By embracing these quantum principles, we can develop a more nuanced, flexible, and effective approach to strategic leadership in our rapidly changing world.

𝐑𝐮𝐧 𝐲𝐨𝐮𝐫 𝐀𝐈 𝐟𝐞𝐚𝐭𝐮𝐫𝐞 𝐭𝐰𝐢𝐜𝐞 𝐨𝐧 𝐭𝐡𝐞 𝐬𝐚𝐦𝐞 𝐢𝐧𝐩𝐮𝐭. 𝐈𝐟 𝐚 𝐝𝐢𝐟𝐟𝐞𝐫𝐞𝐧𝐭 𝐚𝐧𝐬𝐰𝐞𝐫 𝐰𝐨𝐮𝐥𝐝 𝐛𝐞 𝐚 𝐩𝐫𝐨𝐛𝐥𝐞𝐦, 𝐲𝐨𝐮'𝐯𝐞 𝐟𝐨𝐮𝐧𝐝 𝐚 𝐝𝐞𝐭𝐞𝐫𝐦𝐢𝐧𝐢𝐬𝐭𝐢𝐜 𝐫𝐞𝐪𝐮𝐢𝐫𝐞𝐦𝐞𝐧𝐭. It sounds trivial. It's the fastest way I know to tell whether you're about to put AI in the wrong place. Non-deterministic means the same input can produce different output: an LLM, most ML scoring. Deterministic means same input, same output, every time: SQL, a function, a rules engine. Run-it-twice sorts the work into two piles: - A different answer is fine, even useful. Brainstorming, drafting, summaries, fuzzy matching with a human in the loop. A non-deterministic model fits. - A different answer is a problem. Your month-end total, a compliance figure, a data transform, anything you have to reproduce or audit. You need a deterministic core. The trap I see most often is a model doing the second kind of job. An agent that computes the figure directly. A "smart" pipeline that re-decides its schema every run. You can't unit-test a vibe, and "the model said so" won't satisfy a regulator. So keep the deterministic core deterministic, and let the model build it. The LLM writes the SQL, you review it, the SQL returns the same answer forever. The model is excellent at generating the tool; running it in production is the tool's job. Run it twice before you ship. If the two answers have to match, put a deterministic engine at the center and keep the model on the edges. 𝐖𝐡𝐞𝐫𝐞 𝐢𝐧 𝐲𝐨𝐮𝐫 𝐬𝐭𝐚𝐜𝐤 𝐢𝐬 𝐚 𝐧𝐨𝐧-𝐝𝐞𝐭𝐞𝐫𝐦𝐢𝐧𝐢𝐬𝐭𝐢𝐜 𝐦𝐨𝐝𝐞𝐥 𝐝𝐨𝐢𝐧𝐠 𝐚 𝐣𝐨𝐛 𝐭𝐡𝐚𝐭 𝐧𝐞𝐞𝐝𝐬 𝐭𝐡𝐞 𝐬𝐚𝐦𝐞 𝐚𝐧𝐬𝐰𝐞𝐫 𝐞𝐯𝐞𝐫𝐲 𝐭𝐢𝐦𝐞?