About
About
Views are my own and do not represent those of my employer. I share publicly…
Views are my own and do not represent those of my employer. I share publicly…
Articles by Dileep
Activity
22K followers
-
Dileep Pandiya shared this𝗧𝗵𝗲 𝗴𝗮𝗽 𝗯𝗲𝘁𝘄𝗲𝗲𝗻 "𝘁𝗵𝗲 𝗱𝗲𝗺𝗼 𝘄𝗼𝗿𝗸𝗲𝗱" 𝗮𝗻𝗱 "𝗶𝘁'𝘀 𝗿𝗲𝗮𝗱𝘆 𝗳𝗼𝗿 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻" 𝗶𝘀 𝗯𝗶𝗴𝗴𝗲𝗿 𝘁𝗵𝗮𝗻 𝗺𝗼𝘀𝘁 𝘁𝗲𝗮𝗺𝘀 𝘁𝗵𝗶𝗻𝗸. Same five reasons every time. 𝟭. 𝗜𝗴𝗻𝗼𝗿𝗶𝗻𝗴 𝘀𝘁𝗼𝗽 𝗿𝗲𝗮𝘀𝗼𝗻𝘀 • The agent keeps going no matter what the model actually returned • Hit a token limit, needs to call a tool, or genuinely finished, it treats all three the same way • Each one needs its own next step, not a blind continue 𝟮. 𝗢𝗻𝗲 𝗮𝗴𝗲𝗻𝘁, 𝟱𝟬+ 𝘁𝗼𝗼𝗹𝘀 • Feels efficient on paper • In practice the agent starts guessing which tool to use instead of reasoning about it • Splitting into smaller specialists with a tool or two each fixes this almost every time 𝟯. 𝗙𝘂𝗹𝗹 𝗰𝗼𝗻𝘃𝗲𝗿𝘀𝗮𝘁𝗶𝗼𝗻 𝗵𝗶𝘀𝘁𝗼𝗿𝘆 𝗽𝗮𝘀𝘀𝗲𝗱 𝗯𝗲𝘁𝘄𝗲𝗲𝗻 𝗮𝗴𝗲𝗻𝘁𝘀 • Someone always says "just forward everything, more context can't hurt" • It can. It buries the actual task in noise and drives up token cost for nothing • A structured summary with the goal, key decisions, and current state does more with less 𝟰. 𝗨𝗻𝗳𝗶𝗹𝘁𝗲𝗿𝗲𝗱 𝘁𝗼𝗼𝗹 𝗼𝘂𝘁𝗽𝘂𝘁𝘀 • A database call returns thousands of rows, all of it lands in the context window • Now the agent is reasoning about data it never needed • Retrieve, filter, summarize, then hand it to the model 𝟱. 𝗜𝗻𝘁𝗲𝗿𝗮𝗰𝘁𝗶𝘃𝗲 𝗽𝗿𝗼𝗺𝗽𝘁𝘀 𝗶𝗻𝘀𝗶𝗱𝗲 𝗖𝗜/𝗖𝗗 • The pipeline doesn't fail. It just never finishes • An agent pauses waiting for user input inside a pipeline running unattended at 2am • Nobody's there to answer, so the job just hangs None of this comes down to the model being smart enough. It comes down to whether the system around it has any guardrails. #AIagents #SystemDesign #SoftwareArchitecture #AI
-
Dileep Pandiya shared this𝗪𝗮𝘁𝗰𝗵𝗲𝗱 𝗧𝗵𝗲 𝗢𝗱𝘆𝘀𝘀𝗲𝘆 𝘆𝗲𝘀𝘁𝗲𝗿𝗱𝗮𝘆. What stayed with me was what leadership looks like after the decision is made. In engineering, we celebrate the bold call: 1. Choose the platform 2. Start the migration 3. Move fast on AI adoption 4. Commit to a new architecture But the real leadership test comes later. • Can you own the consequences when execution gets difficult? • Can you adapt when reality changes? • Can you keep people steady when uncertainty increases? The decision is only the beginning. Leadership is tested in how you carry your team through the tradeoffs, setbacks and outcomes that follow. #Leadership #EngineeringLeadership #TechLeadership #AILeadership #SystemDesign
-
Dileep Pandiya shared this𝗡𝗼𝗯𝗼𝗱𝘆'𝘀 𝗳𝗶𝗴𝗵𝘁𝗶𝗻𝗴 𝗼𝘃𝗲𝗿 𝘁𝗵𝗲 𝗺𝗼𝗱𝗲𝗹 𝗮𝗻𝘆𝗺𝗼𝗿𝗲 Anthropic is reportedly paying $6B for Decart AI, a 2 year old company most people have never heard of. Not a model company. An inference company. It's part of a pattern. Nvidia paid $20B for Groq's inference tech in December, put $150M into inference startup Baseten in January, and Decart itself went from a $4B valuation in May to a reported $6 to 7B now, a 50% jump in three months. None of these deals are about a smarter model. All of them are about running the models we already have, cheaper and faster. My take is that we spent 2024 and 2025 fighting over the model layer, things like attention mechanisms, context windows, benchmark leaderboards. That layer is commoditized now. Claude, GPT, and Gemini swap the top spot every few weeks. The real margin left is one layer down in batching, caching and hardware aware scheduling. The detail that convinces me this is structural and Decart's team is reportedly joining Anthropic's inference and performance org. That tells me this is a company buying efficiency. If you're building agentic systems that chain multiple LLM calls your cost and latency are being set more by your serving layer than by which model's logo is on the API call. Model choice is becoming the smaller lever. Where is your team spending more engineering time right now, picking the model or optimizing what runs underneath it?
-
Dileep Pandiya shared thisI shipped a tool today that most teams don't think they need. Until it burns them. If you build with Claude, you've written a SKILL.md file. And you've almost certainly shipped it based on gut feel. Here's the problem: skills are just system prompt injections. Nobody is actually measuring if they help. Most teams "evaluate" a skill by eyeballing a few responses. That's not evaluation. That's vibes. And vibes break down fast: ✅ Claude judging its own outputs favors whichever one it sees first ✅ Model updates and skill edits silently regress performance over time ✅ Every engineer scores "good" differently so the label means nothing across a team So I built skilleval. ✅ Blind A/B testing — the skill vs. no skill, run concurrently on every task ✅ Randomized judge — removes position bias entirely, no more "Output A wins because it went first" ✅ Deterministic assertions — hard checks like must_contain / must_not_contain, no LLM opinion needed for the obvious stuff ✅ Regression detection — skilleval diff shows the exact run where a skill started getting worse ✅ Version comparison — test v1 vs v2 head-to-head instead of guessing which is better ✅ CI integration — one GitHub Action blocks any PR that quietly breaks a skill before it ships This is the same discipline we already apply to code. Skills deserve the same rigor. Open source. MIT licensed. Real gaps documented in the repo, I'm not pretending it's finished. If you're running Claude skills in production, how are you evaluating them today? 👉 https://lnkd.in/gZXTfkKN #AI #ClaudeAI #LLM #EngineeringLeadership #OpenSource #Skill
-
Dileep Pandiya shared thisI am happy to announce that I have signed a contract with Apress to publish my upcoming book: Vibecoding: The Future of AI Communication! 🚀 Over the last few years, we have seen the shift from traditional software development to AI-assisted coding. But building functional code isn't enough anymore, we need systems that understand human context, nuance and intent. That's what "vibecoding" is all about: the architecture, data pipelines, and interaction frameworks required for human-centric, multimodal AI. This book is a deep dive into the engineering side of conversational AI, bridging the gap between UX design and agentic system orchestration. It’s designed to be a practical handbook for AI architects, engineers, and product managers. A huge thank you to my brilliant co-authors for embarking on this journey with me and to the entire team at Apress for believing in this vision. We’re deep into drafting and will be sharing more updates, including code templates, data pipelines and a GitHub repository as we go. (The cover below is just a placeholder, stay tuned for the official Apress cover reveal in the coming months!) What’s the biggest challenge you’re currently seeing in multimodal AI and agent orchestration? Let me know below, it might just make it into a chapter! 👇 #Vibecoding #ConversationalAI #ArtificialIntelligence #AgenticSystems #TechBooks #Apress #SoftwareEngineering #AI
-
Dileep Pandiya shared this🧵 I gave a talk at DeveloperWeek last week that made a few architects uncomfortable. Good. 🎯 "Your API Layer Just Became Optional." For 15 years, the API has been the non-negotiable contract between systems. That contract is being renegotiated and no one asked your permission. 📊 Gartner projects 40% of enterprise applications will embed AI agents by end of 2026, up from 5% in 2025. Those agents don't always call your APIs. Here is how they operate instead: 🔍 MCP servers that discover capabilities dynamically 🖥️ Computer use models that operate UIs directly, never touching your endpoints 🤝 A2A protocols that let agents delegate to other agents without a single REST call in the chain Most architecture teams haven't started this conversation yet. My session was that conversation. The session walked through: ✅ Where REST and GraphQL still win 🔄 Where MCP replaces them ⚡ Where computer use agents bypass both entirely All with a concrete decision framework for building systems that serve humans, agents, and everything in between. 🙌 Got to connect with some incredible leaders and have real conversations about the AI integration challenges teams are facing right now. That energy is exactly why I keep showing up to these events. If you haven't mapped how agents interact with your systems yet, that's the conversation to start this week. #APIArchitecture #AIAgents #MCP #DeveloperWeek #SystemDesign #A2A
-
Dileep Pandiya shared thisToday I had the honor of speaking at #LDX3London talking about Lessons from 100 P0 Incidents, 20 years, 5 companies and more outages than I would like to admit. Every P0 has two timelines. One is technical. The other is organizational. Technically similar incidents end very differently and the difference is almost always decided before the pager fires. After 100 of these, the five things I would tell any engineering team to fix before their next P0: 1. Make rollback faster than debugging 2. Add one business metric, not CPU, not memory but did the user actually get what they came for? 3. Write your incident commander runbook before you need it. 4. Before any big change: can you roll it back cleanly and will you know within minutes if it goes wrong? 5. Check your last postmortem, are the action items actually owned and in a sprint slot? The best engineering teams are not the ones that never fail. They are the ones that know how to fail well. Huge thank you to the LeadDev team for having me. The energy in the room was brilliant. 🙏 #LeadDev #Engineering #SRE #IncidentResponse #Reliability #PlatformEngineering #SoftwareEngineering
-
Dileep Pandiya shared this𝗧𝗵𝗲 𝗼𝗻𝗹𝘆 𝗲𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴 𝗺𝗼𝗱𝗲𝗹 𝗰𝗵𝗲𝗮𝘁 𝘀𝗵𝗲𝗲𝘁 𝘆𝗼𝘂'𝗹𝗹 𝗻𝗲𝗲𝗱 𝗶𝗻 𝟮𝟬𝟮𝟲. Most teams pick embedding models based on name recognition. That's why their RAG pipelines underperform. I spent time benchmarking the entire 2026 embedding model landscape so you don't have to. Here's what the data actually says 👇 ☁️ 𝗠𝗮𝗻𝗮𝗴𝗲𝗱 𝗔𝗣𝗜𝘀 : when you need reliability with zero infra: → Gemini-001 leads at MTEB 68.3. Best if you're on GCP. → Voyage-3 outperforms OpenAI by 9.74% across 100 real-world datasets. → OpenAI TE-3 remains the safest enterprise default at $0.13/1M tokens. 🏆 𝗠𝗮𝘅 𝗔𝗰𝗰𝘂𝗿𝗮𝗰𝘆 (𝗢𝗦𝗦) : when accuracy is non-negotiable: → Qwen3-8B is the #1 model globally right now. MTEB 70.2. Apache 2.0. Free. → NV-Embed-v2 from NVIDIA hits 69.3 using latent attention instead of mean pooling. → E5-Mistral from Microsoft — open weights, high accuracy, production ready. ⚖️ 𝗕𝗮𝗹𝗮𝗻𝗰𝗲𝗱 𝗢𝗦𝗦 : the RAG engineer's sweet spot: → BGE-M3 is the only model doing Dense + Sparse + ColBERT hybrid retrieval in ONE model. → Jina-v3 gives you 8,192 token context — longer than any competitor in its class. → Arctic-Embed from Snowflake beats both Cohere and OpenAI on pure retrieval tasks. 📱 𝗟𝗶𝗴𝗵𝘁𝘄𝗲𝗶𝗴𝗵𝘁 / 𝗘𝗱𝗴𝗲 : when you can't afford 7B parameters: → EmbeddingGemma at 300M params is the best sub-500M model on MTEB. Runs on device. → BGE-Small is the battle-tested production workhorse. MTEB 60.7. Never lets you down. → Nomic-v1 has 8k token context at a fraction of the size of LLM-based models. 🔬 𝗗𝗼𝗺𝗮𝗶𝗻 𝗦𝗽𝗲𝗰𝗶𝗳𝗶𝗰 : when general models hit a ceiling: → Nomic Code for developer tools and repository search. → BGE-Finance for SEC filings, risk analysis, and market data retrieval. → BioLORD for clinical notes, PubMed, and biomedical research pipelines. The uncomfortable truth? An open-source model (Qwen3-8B, MTEB: 70.2) is beating every paid API right now. Most teams are paying for embeddings they don't need to. The embedding layer is the first place your retrieval quality breaks — and the last place most engineers look. Save this post. Your future self debugging a RAG pipeline will thank you. ♻️ Repost if this helps your team. 💬 Which embedding model are you using in production? Drop it comments 👇 #AI #MachineLearning #RAG #LLM #NLP #VectorSearch #DataEngineering #MLOps #OpenSource #SoftwareEngineering
-
Dileep Pandiya shared this𝗬𝗼𝘂𝗿 𝗥𝗔𝗚 𝘀𝘆𝘀𝘁𝗲𝗺 𝗱𝗼𝗲𝘀𝗻'𝘁 𝗳𝗮𝗶𝗹 𝗯𝗲𝗰𝗮𝘂𝘀𝗲 𝘆𝗼𝘂 𝗽𝗶𝗰𝗸𝗲𝗱 𝘁𝗵𝗲 𝘄𝗿𝗼𝗻𝗴 𝗺𝗼𝗱𝗲𝗹. Teams spend weeks debating GPT-4 vs Claude, then ship systems that give wrong answers half the time. The problem isn't intelligence. It's information architecture. We tend to focus only on the "RAG" bubble or the "Prompt Engineering" bubble. But the real game is in the center: Context Engineering. It's the discipline of managing the limited workspace (context window) where your model actually lives. Three architectural shifts that move the needle: 𝟭. 𝗛𝗼𝘄 𝘆𝗼𝘂 𝘀𝗽𝗹𝗶𝘁 𝗱𝗼𝗰𝘂𝗺𝗲𝗻𝘁𝘀 𝗰𝗵𝗮𝗻𝗴𝗲𝘀 𝗲𝘃𝗲𝗿𝘆𝘁𝗵𝗶𝗻𝗴 Small chunks surface easily in searches but lack meaning. Large chunks carry context but become unfindable. Most systems default to arbitrary 500-token splits with overlap. This works until you need precision. Better approaches: split on semantic boundaries for research papers, follow document structure for wikis, or embed entire documents first then extract chunks. Match your strategy to your content type. 𝟮. 𝗜𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝘁 𝘀𝘆𝘀𝘁𝗲𝗺𝘀 𝗻𝗲𝗲𝗱 𝗰𝗼𝗼𝗿𝗱𝗶𝗻𝗮𝘁𝗶𝗼𝗻 𝗹𝗮𝘆𝗲𝗿𝘀 Build a decision-making component that evaluates what it knows, identifies gaps, selects the right tools, and adjusts when things fail. This isn't about autonomous agents doing everything. You need something that reformulates bad queries, routes to the right data source, and compresses conversation history when it bloats. 𝟯. 𝗜𝗻𝗳𝗼𝗿𝗺𝗮𝘁𝗶𝗼𝗻 𝗾𝘂𝗮𝗹𝗶𝘁𝘆 𝗱𝗲𝗴𝗿𝗮𝗱𝗲𝘀 𝗼𝘃𝗲𝗿 𝘁𝗶𝗺𝗲 Your system injects bad data that gets reused. It accumulates history until it stops thinking clearly. It stumbles over contradictory sources. Build active cleanup: verify what you retrieve, compress old conversations, prune stale information, store details externally instead of inline. 𝗪𝗵𝗮𝘁 𝘁𝗵𝗶𝘀 𝗹𝗼𝗼𝗸𝘀 𝗹𝗶𝗸𝗲 𝗶𝗻 𝗽𝗿𝗮𝗰𝘁𝗶𝗰𝗲: Rewrite unclear queries. Adjust retrieval strategies in real-time. Compress conversation history to preserve space. Execute tools based on decision logic, not guesswork. This is not prompt engineering. This is systems design for how information reaches your model. The gap between demo and production isn't model capability. It depends on how well you manage what information the model sees and when. Chunking strategy? Context management? Tool orchestration? What breaks first when you scale? #SystemDesign #LLM #RAG #SoftwareEngineering #EngineeringLeadership
Dileep Pandiya commented on a post
4w
CFBR
-
Dileep Pandiya liked thisDileep Pandiya liked thisWe’ve all seen some ridiculously good GTM workflows built in Claude this year 👀 The kind where you dump in CRM notes, call transcripts, emails, account data and it generates a near-perfect account plan: who matters, what changed, what to do next, and what to say. They’re very cool. There’s just one problem: the workflow has a user base of one. Usually the GTM engineer who built it. Everyone else is still digging through Salesforce, piecing together context, and starting from scratch. Agent Teams changes that. It puts agents in every rep's day. 👇👇 ☑️ Inbound lead? Research, qualify, route, sequence. ☑️ Deal just closed won? Find lookalikes, prioritize ones worth pursuing, draft outreach. ☑️ Champion changes jobs? Rebuild the relationship, find the owner, draft the outreach. ☑️ Renewal looking bleak? Pull the history, build the save plan, brief the account owner. The reliability of a workflow, with the judgment of a team. Full context, end to end. Available to everyone, not just the person who built it. Meet ZoomInfo Agent Teams: https://lnkd.in/eEuVivSH
-
Dileep Pandiya liked thisDileep Pandiya liked thisBig news: ZoomInfo has acquired DoubleO.ai. 🚀 Everyone's built an AI agent demo. Almost no one trusts one to run a real GTM motion. Agents hallucinate, behave differently every run, and need a human babysitting every step. DoubleO.ai's orchestration layer fixes that, making multi-agent automation reliable enough to run complete plays end to end. Paired with ZoomInfo's GTM Context Graph (100M+ companies, 500M+ contacts, billions of signals), agents can now: ✅ Track champions who change jobs and engage them at their new company ✅ Turn closed-won deals into new pipeline by finding lookalikes and reaching out with tailored messaging Demos are easy. Reliable GTM execution is the unlock. Read more: https://okt.to/aE73Zr
-
Dileep Pandiya liked thisDileep Pandiya liked thisToday, we're announcing that ZoomInfo has acquired DoubleO.ai, an agentic orchestration platform built for reliable, multi-agent automation. Agents break down the moment they're asked to do more than answer a question. They hallucinate, run the task differently every time, and lose context. The result is impressive demos that no one trusts with a real workflow. Pairing DoubleO.ai's orchestration with ZoomInfo's GTM Context Graph (100M+ companies, 500M+ contacts, and billions of signals) means agents can run complete plays, from signal to action. What that looks like: → Champion tracking: When a champion changes jobs, agents check ICP fit, map the buying committee, and reach out based on your history with them. → Closed-won to new pipeline: When a deal closes, agents learn why the customer bought, draft the case study, find lookalike accounts, and engage each with tailored messaging. DoubleO.ai already powers GTM workflows at Fortune 500 companies and fast-growing software firms. Now it's joining the platform with the context to take it further. Welcome to ZoomInfo, DoubleO.ai team! 🎉 Read the full announcement: https://okt.to/jYzaUp
-
Dileep Pandiya liked thisDileep Pandiya liked thisZoomInfo’s data team is looking for a Sr. Software Engineer to drive the direction of our data services and build world-class big data pipelines from scratch! In this role, you'll leverage your 7+ years of experience in distributed systems, big data frameworks (Spark, Hadoop), and the JVM stack (Java, Scala) to process massive datasets and collaborate closely with data scientists on ML workflows. If you're passionate about large-scale data architectures and want to make technographics a first-class offering—let’s talk. 👉 Drop your application below or reach out directly to join us! #ZoomInfo #Hiring #SoftwareEngineer #BigData #DataPipelines #Spark #Java #Scala #TechJobs #Careers
-
Dileep Pandiya liked thisI am proud of my son Noah Boudreau and all that his startup company, Smart Escape, has accomplished!Dileep Pandiya liked thisWe are happy to announce that the Smart Escape Team has completed another project at the Henry P. Clough Elementary School in Mendon, MA! The system is now active and ready to guide the teachers, students, and faculty to safety away from a fire and out of the building. Huge thanks to the Mendon Fire Department, The Mendon-Upton Regional School District, and the Smart Escape Team for making this all possible. If you are interested in learning more about Smart Escape or interested in seeing the system in action, please visit www.smartescapeusa.com to contact us and schedule a time!
-
Dileep Pandiya liked thisDileep Pandiya liked thisI’m excited to be building out our avionics software leadership team, and I’m looking for a Software Chief Architect to join us. This is an opportunity for a strong technical leader to shape software architecture and engineering across some of the most important and challenging areas of safety-critical avionics. If you have the experience of architecting and certifying safety critical software, then I’d love to connect. The role is open in Long Beach, CA, Plano, TX, and Mesa, AZ.
-
Dileep Pandiya liked thisDileep Pandiya liked this20+ years in tech. Amazon. JPMorgan. And one lesson has stayed with me: Senior engineering is never just about writing better code. Over the years, I’ve designed systems, led engineering teams, worked on large-scale technology programmes, interviewed engineers, mentored professionals, and made decisions where architecture was only one part of the challenge. The other part was always human. How do you explain a difficult trade-off? How do you influence stakeholders? How do you handle ambiguity, conflict, ownership, and risk? How do you communicate like a senior engineer when there is no perfect answer? That is exactly why Alok Sharan and I wrote “System Design and Behavioral Intelligence.” We wanted to bring technical thinking and behavioral intelligence together in one practical guide. Inside, we cover distributed systems, scalability, databases, security, observability, architecture patterns, and complete designs including Slack-like messaging, real-time bidding, distributed key-value stores, and Agentic AI systems. But we also go deeper into STAR+R, leadership storytelling, DORA, RACI, stakeholder management, risk thinking, and how senior engineers are evaluated beyond technical correctness. Because the next level of your engineering career requires more than knowing how systems work. It requires showing how you think, decide, communicate, and lead. 📘 System Design and Behavioral Intelligence If you are preparing for senior engineering, architecture, or Big Tech roles, this is the book we wished we had earlier in our careers. P.S. Preparing for a software engineering role? Check out my book, System Design and Behavioral Intelligence. It will help you: • Build strong System Design foundations and approach design problems with confidence • Develop practical frameworks to navigate complex behavioral and leadership challenges Get Your Copy anywhere internationally: US: https://lnkd.in/eVy2ACxQ UK: https://lnkd.in/evzmHKm4 India : https://lnkd.in/e25PthqU
-
Dileep Pandiya liked thisDileep Pandiya liked thisMeet the latest model in our new Claude 5.5 family. Build your first app, finish the side project, or take on the big code migration.
-
Dileep Pandiya liked thisDileep Pandiya liked thisOur engineering team at ZoomInfo has been absolutely COOKING over the last year.🍳🍳 Talent Autopilot is our new AI recruiting agent. -->Tell it who you need. It finds candidates you might NEVER find through traditional search. It does the work. It finds them, tells you why they fit, and helps start the conversation. -->It runs on ZoomInfo data. The same verified professional data and context that GTM teams rely on every day. We’ve spent 19 years helping companies find their next customer. Now they can use agents to find their next great hire. If you're in recruiting or talent acquistion - check it out here: https://lnkd.in/dC3XdqxS
Experience & Education
Licenses & Certifications
Volunteer Experience
Projects
-
AI Alignment Liaison
The AI Alignment Liaison is a open source generative AI application to advise on how to align AI development processes (within a team or organization) to a set of human values.
Languages
-
English
-
-
Hindi
-
Recommendations received
-
LinkedIn User
-
LinkedIn User
8 people have recommended Dileep
Join now to viewOther similar profiles
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content