Product Value Creation

Few Lessons from Deploying and Using LLMs in Production Deploying LLMs can feel like hiring a hyperactive genius intern—they dazzle users while potentially draining your API budget. Here are some insights I’ve gathered: 1. “Cheap” is a Lie You Tell Yourself: Cloud costs per call may seem low, but the overall expense of an LLM-based system can skyrocket. Fixes: - Cache repetitive queries: Users ask the same thing at least 100x/day - Gatekeep: Use cheap classifiers (BERT) to filter “easy” requests. Let LLMs handle only the complex 10% and your current systems handle the remaining 90%. - Quantize your models: Shrink LLMs to run on cheaper hardware without massive accuracy drops - Asynchronously build your caches — Pre-generate common responses before they’re requested or gracefully fail the first time a query comes and cache for the next time. 2. Guard Against Model Hallucinations: Sometimes, models express answers with such confidence that distinguishing fact from fiction becomes challenging, even for human reviewers. Fixes: - Use RAG - Just a fancy way of saying to provide your model the knowledge it requires in the prompt itself by querying some database based on semantic matches with the query. - Guardrails: Validate outputs using regex or cross-encoders to establish a clear decision boundary between the query and the LLM’s response. 3. The best LLM is often a discriminative model: You don’t always need a full LLM. Consider knowledge distillation: use a large LLM to label your data and then train a smaller, discriminative model that performs similarly at a much lower cost. 4. It's not about the model, it is about the data on which it is trained: A smaller LLM might struggle with specialized domain data—that’s normal. Fine-tune your model on your specific data set by starting with parameter-efficient methods (like LoRA or Adapters) and using synthetic data generation to bootstrap training. 5. Prompts are the new Features: Prompts are the new features in your system. Version them, run A/B tests, and continuously refine using online experiments. Consider bandit algorithms to automatically promote the best-performing variants. What do you think? Have I missed anything? I’d love to hear your “I survived LLM prod” stories in the comments!

There’s a fine line between saying “no” because of attitude and saying “no” because you understand the value of what you bring to the table. Early on, I realized it wasn’t about followers, views, or appearances. It was about attention to detail, the process and the standards I had set for myself and my team. When clients asked to lower rates or push budgets, the response was simple: That’s my price. No over-explaining, no defending, no justifying. Confidence in the value you create is often more persuasive than any argument about experience or past projects. This mindset helps attract the right clients as well. The people who value your approach and respect your standards naturally gravitate toward working with you. And sometimes, it allows you to say “no” to opportunities that don’t align, preserving focus, quality and integrity. It’s also about presence. In client interactions, nothing replaces direct engagement. Even with a capable team, certain conversations, especially first calls or high-stakes projects, benefit from your direct involvement. People want to feel the commitment, the clarity, & the vision firsthand. That connection often determines whether a client signs on or walks away. At the end of the day, value is in how you position it, how you communicate it & how you stand by it. The right clients recognize that and the wrong ones fade away. And that’s exactly how you build sustainable, meaningful work that makes a real impact. #graphicdesign

Most people still think of LLMs as “just a model.” But if you’ve ever shipped one in production, you know it’s not that simple. Behind every performant LLM system, there’s a stack of decisions, about pretraining, fine-tuning, inference, evaluation, and application-specific tradeoffs. This diagram captures it well: LLMs aren’t one-dimensional. They’re systems. And each dimension introduces new failure points or optimization levers. Let’s break it down: 🧠 Pre-Training Start with modality. → Text-only models like LLaMA, UL2, PaLM have predictable inductive biases. → Multimodal ones like GPT-4, Gemini, and LaVIN introduce more complex token fusion, grounding challenges, and cross-modal alignment issues. Understanding the data diet matters just as much as parameter count. 🛠 Fine-Tuning This is where most teams underestimate complexity: → PEFT strategies like LoRA and Prefix Tuning help with parameter efficiency, but can behave differently under distribution shift. → Alignment techniques- RLHF, DPO, RAFT, aren’t interchangeable. They encode different human preference priors. → Quantization and pruning decisions will directly impact latency, memory usage, and downstream behavior. ⚡️ Efficiency Inference optimization is still underexplored. Techniques like dynamic prompt caching, paged attention, speculative decoding, and batch streaming make the difference between real-time and unusable. The infra layer is where GenAI products often break. 📏 Evaluation One benchmark doesn’t cut it. You need a full matrix: → NLG (summarization, completion), NLU (classification, reasoning), → alignment tests (honesty, helpfulness, safety), → dataset quality, and → cost breakdowns across training + inference + memory. Evaluation isn’t just a model task, it’s a systems-level concern. 🧾 Inference & Prompting Multi-turn prompts, CoT, ToT, ICL, all behave differently under different sampling strategies and context lengths. Prompting isn’t trivial anymore. It’s an orchestration layer in itself. Whether you’re building for legal, education, robotics, or finance, the “general-purpose” tag doesn’t hold. Every domain has its own retrieval, grounding, and reasoning constraints. ------- Follow me (Aishwarya Srinivasan) for more AI insight and subscribe to my Substack to find more in-depth blogs and weekly updates in AI: https://lnkd.in/dpBNr6Jg

I found this meme funny… but also strikingly accurate. Many CEOs are rushing into AI with huge enthusiasm, but often without clarity on what specific problem they’re solving. The result? Exactly what you see here. After 3+ years partnering with companies on conversational AI solutions, I’ve seen this pattern repeat countless times. Organizations invest in AI, then wonder why they’re not seeing ROI. The real challenge isn’t “Do we need AI?” (we do). It’s “How do we implement it to create measurable, sustainable value?” Here’s what I’ve learned separates successful AI implementations from expensive experiments: Start with the problem, not the technology – Define outcomes before choosing tools. Establish clear success metrics – If you can’t measure it, you can’t improve it Align strategy across stakeholders – Technical teams and business leaders must speak the same language. Focus on value, not features – Shiny doesn’t always mean useful The technology is ready. What’s often missing is the strategic bridge between business objectives and technical execution. I’ve worked with CTOs who knew exactly what they wanted to build but couldn’t quantify business impact. I’ve advised executives who had clear ROI targets but no technical roadmap. The magic happens when strategy and execution align. What’s been your experience with AI implementation? Are you seeing real value — or just expensive experiments? #AI #ConversationalAI #DigitalTransformation #BusinessStrategy #TechLeadership

I recently spent time getting more hands-on with LLM & Agentic AI engineering through Ed Donner's training. Instead of stopping at examples, I built a mini multi-agent logistics delivery optimization framework. Building real AI systems quickly makes one thing clear: 𝙏𝙝𝙚 𝙝𝙖𝙧𝙙 𝙥𝙖𝙧𝙩 𝙞𝙨𝙣’𝙩 𝙩𝙝𝙚 𝙢𝙤𝙙𝙚𝙡 — 𝙞𝙩’𝙨 𝙩𝙝𝙚 𝙖𝙧𝙘𝙝𝙞𝙩𝙚𝙘𝙩𝙪𝙧𝙚 𝙙𝙚𝙘𝙞𝙨𝙞𝙤𝙣𝙨 𝙖𝙧𝙤𝙪𝙣𝙙 𝙞𝙩. A few practical lessons: 1. 𝗟𝗟𝗠 𝗺𝗼𝗱𝗲𝗹 𝘀𝗲𝗹𝗲𝗰𝘁𝗶𝗼𝗻 𝗶𝘀 𝗳𝗮𝗿 𝗺𝗼𝗿𝗲 𝗻𝘂𝗮𝗻𝗰𝗲𝗱 𝘁𝗵𝗮𝗻 𝗰𝗼𝘀𝘁 𝘃𝘀 𝗹𝗮𝘁𝗲𝗻𝗰𝘆. Trade-offs: • reasoning maturity for complex planning • context window & memory strategy • proprietary models vs smaller open models • infra costs (GPU/hosting) vs token-based API costs • tool-calling reliability & structured output adherence • benchmark performance vs real task behavior • model stability across releases In practice, it becomes a hybrid strategy: 𝘀𝗺𝗮𝗹𝗹𝗲𝗿/𝗰𝗵𝗲𝗮𝗽𝗲𝗿 𝗺𝗼𝗱𝗲𝗹𝘀 𝗳𝗼𝗿 𝗿𝗼𝘂𝘁𝗶𝗻𝗲 𝘁𝗮𝘀𝗸𝘀 + 𝗦𝗟𝗠 𝘄𝗶𝘁𝗵 𝗳𝗶𝗻𝗲-𝘁𝘂𝗻𝗶𝗻𝗴 𝗳𝗼𝗿 𝗱𝗼𝗺𝗮𝗶𝗻 𝗽𝗿𝗼𝗯𝗹𝗲𝗺𝘀 + 𝘀𝘁𝗿𝗼𝗻𝗴𝗲𝗿 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 𝗺𝗼𝗱𝗲𝗹𝘀 𝗳𝗼𝗿 𝗰𝗼𝗺𝗽𝗹𝗲𝘅 𝗱𝗲𝗰𝗶𝘀𝗶𝗼𝗻𝘀. 𝟮. 𝗗𝗲𝘃𝗲𝗹𝗼𝗽𝗺𝗲𝗻𝘁 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 𝗮𝘀 𝗺𝘂𝗰𝗵 𝗮𝘀 𝘁𝗵𝗲 𝗟𝗟𝗠: Many AI demos over-engineer the stack. In reality, simplicity, latency, security and reliability matter more than novelty. • Use orchestration frameworks only where coordination complexity exists • Combine prompts with structured outputs to reduce ambiguity • Watch serialization and tool-call overhead — they impact latency and UX • Reduce unnecessary LLM calls when deterministic code can solve the task Besides lowering token cost, this improves context efficiency, letting models focus on real reasoning. Sometimes best architecture decision is 𝙣𝙤𝙩 𝙞𝙣𝙩𝙧𝙤𝙙𝙪𝙘𝙞𝙣𝙜 𝙖𝙣𝙤𝙩𝙝𝙚𝙧 𝙡𝙖𝙮𝙚𝙧. 3. 𝗕𝗶𝗴𝗴𝗲𝗿 𝗺𝗼𝗱𝗲𝗹𝘀 ≠ 𝗯𝗲𝘁𝘁𝗲𝗿 𝗼𝘂𝘁𝗰𝗼𝗺𝗲𝘀 Smaller models with fine-tuning on domain data can perform more consistently than larger ones. Fine-tuning helps when: • tasks are repetitive but require precision • domain vocabulary is specialized • prompts become fragile But 𝗳𝗶𝗻𝗲-𝘁𝘂𝗻𝗶𝗻𝗴 𝗮𝗹𝘀𝗼 𝗶𝗻𝘁𝗿𝗼𝗱𝘂𝗰𝗲𝘀 𝗹𝗶𝗳𝗲𝗰𝘆𝗰𝗹𝗲 𝗼𝘃𝗲𝗿𝗵𝗲𝗮𝗱. Base model upgrades trigger retesting and partial rewrites. 4. 𝗧𝗵𝗲 𝗿𝗲𝗮𝗹 𝗴𝗮𝗽: 𝗽𝗿𝗼𝘁𝗼𝘁𝘆𝗽𝗲 → 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻 Demos are easy. Production requires 𝙚𝙫𝙖𝙡𝙪𝙖𝙩𝙞𝙤𝙣 𝙛𝙧𝙖𝙢𝙚𝙬𝙤𝙧𝙠𝙨, 𝙤𝙗𝙨𝙚𝙧𝙫𝙖𝙗𝙞𝙡𝙞𝙩𝙮, 𝙨𝙚𝙘𝙪𝙧𝙞𝙩𝙮, 𝙥𝙚𝙧𝙛𝙤𝙧𝙢𝙖𝙣𝙘𝙚, 𝙘𝙤𝙨𝙩 𝙜𝙤𝙫𝙚𝙧𝙣𝙖𝙣𝙘𝙚 & 𝙜𝙪𝙖𝙧𝙙𝙧𝙖𝙞𝙡𝙨. That’s where most engineering effort goes. 𝟱. 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝗳𝗼𝗿 𝗹𝗲𝗮𝗱𝗲𝗿𝘀 𝗿𝘂𝗻𝗻𝗶𝗻𝗴 𝗔𝗜 𝗽𝗿𝗼𝗴𝗿𝗮𝗺𝘀 Many AI conversations focus on SDLC productivity- Useful but the bigger opportunity is 𝙧𝙚𝙞𝙢𝙖𝙜𝙞𝙣𝙞𝙣𝙜 𝙡𝙚𝙜𝙖𝙘𝙮 𝙗𝙪𝙨 𝙥𝙧𝙤𝙘𝙚𝙨𝙨𝙚𝙨 𝙪𝙨𝙞𝙣𝙜 𝘼𝙜𝙚𝙣𝙩𝙞𝙘 AI. By simply automating existing steps, we risk making inefficient tasks efficient and missing the real transformation.

Steve Jobs once observed that the disease of big companies is their ability to confuse process for content. He warned that organizations eventually favor process because more people excel at process than content creation, leading companies to become fixated on the means rather than the ends. Process is the "how" — the frameworks, meetings, documentation, workflows, and operational mechanics of getting work done. Content is the "what" — the actual products, features, and experiences that deliver value to customers. It's the creative output that matters. With the rise of AI, this insight has become more profound and urgent than ever before. AI excels at exactly what our organizations have spent decades optimizing: executing processes, following rules, and automating repetitive tasks. As these capabilities are increasingly handled by AI, what remains uniquely valuable is human creativity, insight, and vision—the very "content" that Jobs spoke about. Yet here's the paradox: Just as human creativity becomes our most critical differentiator, our organizations continue pushing us toward process orientation. Product teams spend their days in roadmap reviews, status updates, and framework applications rather than in creative exploration and customer discovery. We're strengthening the very muscle that AI is rapidly making obsolete while neglecting the creative capacity that makes us irreplaceable. Consider the iPhone. It didn't emerge from a perfect roadmap review or a flawless OKR execution. It came from Jobs' obsession with the content—the experience, the interface, the feeling of holding the internet in your hand. He famously bypassed normal processes, creating a secretive, content-focused team that prioritized the creative vision over established procedures. The most successful AI products aren't emerging from perfect PRD templates or flawlessly executed OKR processes. They're coming from teams that give themselves permission to explore, create, and iterate rapidly—teams that prioritize content over comfort. For AI product leaders, this means: 1. Automating process work. Use AI to handle the processes that consume your creative energy. Let it draft your status reports, summarize meetings, and track metrics so you can focus on the creative work only humans can do. 2. Creating space for genuine creativity. Carve out significant time for exploration, ideation, and customer interaction. Your most valuable contribution isn't managing process—it's discovering the unexpected insights that lead to breakthrough products. 3. Rewarding content over process excellence. In a world where AI can execute processes flawlessly, we need to shift our reward systems toward valuing creative output, novel insights, and customer impact. As AI increasingly handles the how, humans must focus on the what and why. The companies that thrive will be those that use AI to handle process work while unleashing human creativity to focus on content—the true source of value.

It’s going to take some time to see which legal AI company will win the market. New logos and revenue are positive signs, but increased usage & wallet share over time are probably far more relevant. My experience in the ALSP world suggests that the initial buy decision provides a mixed signal. For example: A legal department may take the first step of “buying” an engagement attorney who’s great at spotting legal issues—but ultimately adds little value to the organization. As it turns out, it’s not enough to be able to merely give good legal answers; you have to also be embedded within the organizational context you operate in. Which requires you to be integrated into all kinds of workflows and processes. You have to truly understand the business you’re operating in. Like how the legal team interacts with the sales/business teams. How quickly you have to respond to stakeholders. Do they use CLM? Which one? You have to live in the same communications and collaboration software as everyone else. I believe the same criteria applies to legal AI. For example, in contracting, a given AI tool may be pretty good at flagging problematic clauses in an agreement. But to maximize the value of that information, it has to be embedded within the workflows, communication channels, and broader context of the organization. Narrow legal “excellence” isn’t going to cut it. No matter how good it looks on a demo or in a pilot. From my vantage point it’s incredibly difficult to tell which of these hot new AI startups meet that standard. Do the startups that are getting the most attention provide tech that's truly embedded within their client/customer organizations? I don’t know. My instinct is to say that the best proxy for that isn't new logo and bookings. Instead, it's whether you're seeing increased usage & wallet share from your buyers over time. And that depends on how things play out *after* the initial buy decision.

Trust is the real bottleneck to AI impact, not GPUs or models. I went through the SAS Data and AI Impact Report. It is one of the clearest looks at what actually drives outcomes in the enterprise. Here is the short version. You can also find the complete report here – https://lnkd.in/d7XfVKNM What the report highlights • Generative AI usage is up, and agentic AI is rising, but traditional ML still underpins real production work. • Most teams say they “trust” AI, yet many lack the governance, explainability, and monitoring needed to prove it. That gap lowers ROI. • ROI improves when goals are value focused. Customer experience, growth, resilience, and time to value outperform pure cost cutting. • The biggest blockers are weak data foundations, inconsistent governance, and skills gaps. • Maturity varies by industry, but leaders share the same pattern. Centralized data, accountable governance, and an end to end AI lifecycle. Why this helps enterprises • It gives a benchmark. Use trust and impact indices to see where you stand and where to invest next. • It links trust to hard results. Governance is not a checkbox. It is how you improve returns and reduce surprises. • It focuses on foundations. Good data, clear policy, and lifecycle oversight beat ad hoc pilots. My take • Move from “save cost” to “create value.” Prioritize customer experience, decision speed, and new revenue paths. • Treat trust like an operating system. Build a reusable layer for governance, explainability, bias testing, evaluation, and monitoring. Use it across all use cases. • Prepare for agentic AI with data work first. Consolidate data, define permissions, and track lineage. Agents will only be as good as the operating environment you give them. • Invest in skills. Teach builders evaluation and safety. Teach business teams how to measure decision quality. • Start small, measure fast, scale what works. Make ROI reviews a habit, not a milestone. Why this matters now AI has moved from pilots to core workflows. If trust lags, risk scales faster than value. If trust leads, value compounds. This report offers a practical map for leaders to shift from enthusiasm to impact. If you lead data or AI in your company, block time with your team this week. Align on foundations, governance, and near term value. Then execute. #data #ai #agenticai #sas #theravitshow

For decades, 'legal tech' meant one thing: building complex, expensive software to help big law firms bill more hours, more efficiently. The entire industry was built to serve the lawyer. That era is officially over. The real, multi-trillion dollar opportunity was never about making lawyers slightly more productive, it was about serving the millions of small businesses and individuals who couldn't afford them in the first place. A new wave of startup founders understands that the future isn't about selling software to law firms, but about delivering legal outcomes to everyone else. This shift is happening in real-time so when I met Andrew Guzman at OpenLaw, with a mission of making legal services accessible and on-demand, I was excited to get involved. Their momentum highlights a broader trend we're seeing. Devalued Currency: On-premise enterprise software sold in multi-year contracts to the top 200 law firms. New Currency: On-demand, transparently-priced legal services delivered through a marketplace that empowers both the client and the independent lawyer. Here’s how the next generation of legal tech founders are building: ✔️They Focus on the Client Experience, Not the Lawyer Workflow. The old guard built tools to optimize tasks within a law firm. The next gen are obsessed with the client's journey. They ask: "How can we get a small business a simple, fixed-fee contract review in 24 hours?" This client-centric obsession, rather than lawyer-centric optimization, is the single biggest mindset shift in the industry. ✔️ They Use AI for Access, Not Just Efficiency. First-gen legal tech used AI to help a $1k/hour lawyer find a document 10% faster. The new generation uses AI to automate routine tasks, enabling a marketplace of lawyers to offer services at a price point small businesses can actually afford. AI isn't a tool to enhance the old model, it's a weapon to unlock a completely new market. ✔️ They Sell Predictability First, Legal Services Second. The biggest barrier for a small business isn't a lack of legal documents, it's the paralyzing fear of surprise bills and hiring the wrong expert. Instead the new gen build products that offer fixed-fee packages, transparent reviews and clear project scopes, ensuring a customer knows the exact cost and deliverable upfront. They understand that what they’re really selling is predictability. The future of legal tech doesn't look like a piece of software. It looks like a simple, elegant experience that finally gives businesses and individuals the expert help they really need. A huge congrats to the OpenLaw team for closing $3.5M and leading the charge. Let's go! 🚀 🚀 🚀 The LegalTech Fund, Wisdom Ventures, Mindful Venture Capital, Flint Capital, Slauson & Co., Techstars, Everywhere Ventures