About
For 18 years at Microsoft…
Articles by Imran
Activity
8K followers
-
Imran Siddique shared thisMy time at OPAQUE Systems is coming to a close, and next, I'm looking forward to joining Anthropic. When I joined in June, I had one goal: move the company 𝗳𝗿𝗼𝗺 𝗰𝗼𝗻𝗳𝗶𝗱𝗲𝗻𝘁𝗶𝗮𝗹 𝗔𝗜 𝘁𝗼 𝘃𝗲𝗿𝗶𝗳𝗶𝗮𝗯𝗹𝗲 𝗔𝗜. Confidential computing protects data while it runs. A customer, an auditor or a regulator still cannot see what the agent did with it. Everyone was building agent governance, and almost nobody could prove it was enforced. So that is what we built, in four and a half months: - 𝗢𝗣𝗔𝗤𝗨𝗘 𝟯.𝟬 reached GA on September 30 and put agent controls next to confidential computing in one platform. - 𝗧𝗥𝗔𝗖𝗘, a signed record of what an agent ran and what it did, became a The Linux Foundation project on August 25, developed with AMD, Intel, Microsoft and TII. SecurityWeek, Help Net Security and Infosecurity Magazine covered it. - 𝗖𝗼𝗻𝗳𝗶𝗱𝗲𝗻𝘁𝗶𝗮𝗹 𝗠𝗖𝗣, 𝗰𝗔𝟮𝗔 𝗮𝗻𝗱 𝗪𝗲𝗶𝗴𝗵𝘁 𝗖𝘂𝘀𝘁𝗼𝗱𝘆 𝗠𝗮𝗻𝗶𝗳𝗲𝘀𝘁 took the same proof to tool calls, agent handoffs and model weights. - 92 releases, 10 repositories with an OpenSSF passing badge, 16 research papers with DOIs, and more than 670,000 PyPI downloads since June. 𝟭𝟭𝟵 𝗽𝗲𝗼𝗽𝗹𝗲 𝗯𝗲𝘀𝗶𝗱𝗲𝘀 𝗺𝗲 got code merged. It also put OPAQUE in the rooms where the rules get written. We joined Agentic AI Foundation and Appia Foundation in July and NVIDIA's Open Secure AI Alliance in September, and TRACE is now proposed as an AAIF Sandbox project. I edit Article 12 (record-keeping) in Appia's EU AI Act group, wrote the detection and evidence section of a CoSAI WS4 whitepaper, and ran TRACE live with Joshua Woodruff's ATF and Dick Hardt's AAuth for the Cloud Security Alliance. Aaron Fulkerson and Rishabh Poddar, thank you for the bet. To the engineering team, thank you for picking the hard version every time. To Qiang Xu and every contributor, it is yours too. The open projects keep going at agentrust-io.com, vendor-neutral, and I keep maintaining them. 𝘗𝘳𝘰𝘶𝘥 𝘰𝘧 𝘸𝘩𝘢𝘵 𝘸𝘦 𝘣𝘶𝘪𝘭𝘵 𝘵𝘰𝘨𝘦𝘵𝘩𝘦𝘳.Imran Siddique shared thisAn exciting new chapter for OPAQUE. Today, we’re proud to share a new agreement with Anthropic that carries OPAQUE’s work forward. Ten years ago, OPAQUE began as a research project at UC Berkeley’s RISELab around an idea many thought was impossible. Together, we turned that research into real-world technology, helped define the category of confidential AI, and advanced verifiable AI as critical infrastructure for the AI era. This team proved it was not just possible, but helped set a new standard. Along the way, we worked to advance how secure AI systems are built. We made the world a little better and safer. To our customers, partners, investors and every OPAQUEling: thank you. We can all be proud of what we built together. Read more: https://lnkd.in/g43MSknzFrom Berkeley Research to the Future of AI: OPAQUE’s Next ChapterFrom Berkeley Research to the Future of AI: OPAQUE’s Next Chapter
-
Imran Siddique shared thisA supplier shares pricing data with a retailer for internal forecasting only. The retailer's agent hands part of the job to a second agent, which emails a slice of that data to a vendor. Six months later the supplier asks: was that allowed? To answer from the records, five things have to line up for that one email. The agreement and its version. The authority the agent had through every hand-off. What was running and where. A policy check that happened before the email, not after. And a record somebody who trusts none of us can check. I went through what exists for each of these, expecting holes. 𝗘𝘃𝗲𝗿𝘆 𝗽𝗶𝗲𝗰𝗲 𝗮𝗹𝗿𝗲𝗮𝗱𝘆 𝗵𝗮𝘀 𝗮 𝘀𝘁𝗮𝗻𝗱𝗮𝗿𝗱, 𝗮𝗻𝗱 𝗺𝗼𝘀𝘁 𝗼𝗳 𝘁𝗵𝗲𝗺 𝗮𝗿𝗲 𝗴𝗼𝗼𝗱. Dataspace Protocol and ODRL for the agreement. OAuth and MCP authorization for the first access. The IETF agent token draft for hand-offs. SPIFFE, RATS and EAT for what ran. SCITT, now RFC 9943, for receipts. What none of them does is connect to the others. The agreement says "internal forecasting only" and nothing carries that sentence through the hand-off. The delegation chain checks, but it does not point back to the agreement. 𝗘𝗮𝗰𝗵 𝘀𝘁𝗮𝗻𝗱𝗮𝗿𝗱 𝗮𝗻𝘀𝘄𝗲𝗿𝘀 𝗶𝘁𝘀 𝗼𝘄𝗻 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻 𝗮𝘁 𝗶𝘁𝘀 𝗼𝘄𝗻 𝗱𝗼𝗼𝗿. The part that bothers me most: 𝗮 𝗺𝗶𝘀𝘀𝗶𝗻𝗴 𝗿𝗲𝗰𝗼𝗿𝗱 𝗮𝗻𝗱 𝗮 𝗽𝗮𝘀𝘀𝗶𝗻𝗴 𝗿𝗲𝗰𝗼𝗿𝗱 𝗹𝗼𝗼𝗸 𝘁𝗵𝗲 𝘀𝗮𝗺𝗲 𝘄𝗵𝗲𝗻 𝗻𝗼𝗯𝗼𝗱𝘆 𝗶𝘀 𝗰𝗼𝘂𝗻𝘁𝗶𝗻𝗴. Try it on one protected action from last week. Walk it back to the agreement. If a link were missing, would anything in your stack tell you? The full map, with a link to every standard, is in this week's edition. If you know an adopted profile that already solves this, tell me. #AgentSecurity #Standards
-
Imran Siddique reacted on thisImran Siddique reacted on thisI've said some version of the same sentence in nearly every keynote, podcast and customer meeting: "𝗮𝘀𝘀𝘂𝗺𝗲 𝘆𝗼𝘂𝗿 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁𝘀 𝗰𝗮𝗻 𝗯𝗲 𝗰𝗼𝗺𝗽𝗿𝗼𝗺𝗶𝘀𝗲𝗱, 𝗮𝗻𝗱 𝘁𝗿𝗲𝗮𝘁 𝘁𝗵𝗲𝗺 𝗹𝗶𝗸𝗲 𝗶𝗻𝘀𝗶𝗱𝗲𝗿𝘀." This weekend Satya Nadella said it to the whole industry. He called on companies to assume advanced agentic models can be compromised and to build controls that treat them like trusted insiders. His centerpiece is an emergency brake that lets an authorized human stop a model mid-task. And he draws a hard line between the model that supplies intelligence and the harness that grants authority, so a model can never both act and vouch for its own behavior. https://lnkd.in/gcEt5UWd We've watched example after example of agents with real credentials doing things nobody authorized, at machine speed. The math was never going to be kind. At a 1% chance of a bad action per agent, a thousand agents makes an incident a near-certainty. That's why OPAQUE built enforcement the agent can't touch. Before an agent touches data, cryptographic enforcement proves exactly what it is. Before and during the work, policy is enforced cryptographically. After, a tamper-proof audit trail records what actually happened. Before. During. After. OPAQUE was an OG here because Raluca Ada Popa, Ion Stoica, and Rishabh Poddar built the company on this idea out of UC Berkeley RISElab, years before the market To put this into practice, I've dropped two open-source resources in the comments: a working implementation and sort of study guide for running AI agents without handing them the keys, and AgenTrust, open specifications for verifiable agents. So, for all of you cats who told me, "really this isn't a problem,"...well, I just want y'all to be safe out there. As such, check the first comment for links to my open source implementation for securing personal agents, and the definitive open source framework for Agent Goverance: AgenTrust-io.com.
-
Imran Siddique shared thisKhalil and I have known each other since college. His note about the books in my bag brought back some memories. Knowing that my book is bringing him back to reading on paper means a lot. Thank you, Khalil.Imran Siddique shared thisSome passions never change, and some journeys are truly inspiring! I've known Imran Siddique since our college days, and one thing I vividly remember about him is his love for books. He always had a book in his hand, apart from the ones already tucked away in his bag! Fast forward to today, and it's amazing to see that same book lover become a published author! And here's something special for me too — after almost 15 years, I'm finally picking up a physical book instead of reading on a screen! And what better way to restart this habit than with a book written by someone you've known since college days, on topics you're genuinely passionate about — System Design and Software Architecture! It's the perfect combination of reconnecting with an old habit, revisiting college memories, and exploring something I truly enjoy. Super excited to get my hands on it and dive in. Looking forward to seeing where this reading journey takes me! Kudos, Imran! It's wonderful to see your passion for reading evolve into a passion for writing and sharing knowledge. Wishing you many more milestones ahead! #SystemDesign #SoftwareArchitecture #SystemArchitecture #SoftwareEngineering #BookReading #ReadingJourney #PublishedAuthor #CollegeMemories #TechBooks #ContinuousLearning
-
Imran Siddique reposted thisImran Siddique reposted thisWould you trust an AI request handled by a machine you’ve never seen before? 🤔 That’s one of the harder questions we’ll tackle next week! 🚨 Join us Monday, October 12 at 5 PM ET / 2 PM PT for Mesh-LLM & Trusting a Stranger’s Machine, Part 2: Evidence You Can Check. When an AI request runs on someone else’s machine, simply being told “it worked” isn’t enough. So what if you could actually check the evidence? 🔗 Mesh-LLM’s Evidence plugin records signed, hash-chained evidence on each node and sends the signed records to both sides of the exchange. We’ll walk through a demo showing: → Two-sided confirmation → A paid Lightning exchange → How the exchange can be represented alongside a TRACE record → What additional runtime and hardware evidence can tell us → And, importantly, where the limits of this approach are Speakers: Michael Neale is the founder and creator of Mesh-LLM, an open-source framework focused on connecting and orchestrating AI models and agents. His work explores practical approaches to building more flexible, collaborative AI systems and making multi-model workflows easier to develop and deploy. Steven Mih is the founder of Action State, focused on building infrastructure and tools that make AI agents more reliable, observable, and accountable. His work centers on helping organizations understand and manage what agents do as they operate across increasingly complex systems. Imran Siddique is the founder of agentrust-io and works on TRACE, an open specification for portable, verifiable records of AI agent runs. His work focuses on creating a common way to capture and verify what AI agents do, enabling greater transparency, trust, and portability across agentic systems. If AI agents are going to work across machines we don’t control, evidence matters. Register today and do not miss this amazing opportunity to learn from the biggest in the industry: https://lnkd.in/eNk-AmMr
-
Imran Siddique reposted thisImran Siddique reposted thisMost people think scaling means adding more. More servers. More services. More databases. More layers. More tools. But what if the smartest way to scale… is to subtract? That is the idea that stayed with me after reading Architecting at Scale by Imran Siddique A Microsoft engineering leader with 18 years of experience, Siddique brings something valuable to this subject: practical engineering thinking rather than technology hype. Explore the book here:👇 https://packt.link/ijsD3 And one principle particularly What stood out to me: “Scale by Subtraction.” Before adding another service, ask: What is the actual bottleneck? Before introducing another layer, ask: Can we solve the problem more simply? Before making an architectural decision, ask: What measurable value will this complexity create? And perhaps most importantly: Can we reverse the decision if it doesn't work? This is where the book becomes more than a system-design book. It becomes a lesson in decision-making. The book follows the evolution of ShopFlow, from a startup MVP to a global AI-native platform, and uses that journey to explore the difficult trade-offs behind scaling modern systems. From: ✓ Vertical vs. horizontal scaling ✓ Microservices and service decomposition ✓ Database sharding ✓ Caching and event-driven architecture ✓ Zero Trust and mTLS ✓ Observability and resilience ✓ High availability and capacity planning ✓ FinOps and cost optimisation ✓ AI-first architecture ✓ Autonomous agents and AI-driven automation But the deeper message is simple: Complexity is not proof of sophistication. Sometimes, the most sophisticated architectural decision is knowing what not to build. That thought applies far beyond software. In organisations, we add processes. In education, we add committees. In business, we add meetings. In careers, we add certifications. In technology, we add tools. And sometimes, the real breakthrough comes from asking: “What can I remove?” For me, that is the biggest takeaway from this book. Scale should not mean making everything bigger. It should mean making the right things stronger. A valuable read for software architects, CTOs, engineering leaders, senior developers, and anyone involved in designing systems that need to grow without becoming unnecessarily complicated. As a professor, I also found the underlying philosophy relevant beyond technology: Measure. Simplify. Experiment. Learn. Scale what works. Remove what doesn't. That may be one of the most sustainable ways to build anything. Have you ever solved a complex problem simply by removing something instead of adding something? I would love to hear your experience. 👉 Share this with someone who is building systems, teams or organisations for scale. 👉 Follow Dr. Sanjeev Kumar Sabharwal Packt #SystemDesign #SoftwareArchitecture #Scalability #AI #ArtificialIntelligence #CloudComputing #TechLeadership #EngineeringLeadership #DistributedSystems #Microservices #AIFirst
-
Imran Siddique reposted thisImran Siddique reposted thisMust-Read for 𝗙𝗼𝘂𝗻𝗱𝗲𝗿𝘀 and 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘀! I just wrapped up Part 1 of "𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝗶𝗻𝗴 𝗮𝘁 𝗦𝗰𝗮𝗹𝗲" by Imran Siddique, and it’s an absolute masterclass in technical engineering. Imran brilliantly maps out the transition of a system from MVP into a global, resilient architecture. As the Founder of PRONATIVE AI, two concepts from the opening chapters deeply resonated with me: 1️⃣ 𝗦𝗰𝗮𝗹𝗲 𝗯𝘆 𝗦𝘂𝗯𝘁𝗿𝗮𝗰𝘁𝗶𝗼𝗻 & 𝗧𝗵𝗲 𝗭𝗲𝗿𝗼-𝗦𝗰𝗮𝗹𝗶𝗻𝗴 𝗥𝘂𝗹𝗲: The powerful reminder that scaling isn’t always about adding infrastructure—it's often about removing unnecessary complexity, shedding technical debt, and knowing when architectural inaction is costlier than a cloud tax. 2️⃣ 𝗔𝘃𝗼𝗶𝗱𝗶𝗻𝗴 𝘁𝗵𝗲 "𝗡𝗲𝘁𝗳𝗹𝗶𝘅 𝗘𝗻𝘃𝘆 𝗧𝗿𝗮𝗽": A sharp, realistic take on avoiding resume-driven development and over-engineering before your business logic demands it. Imran cleanly bridges the gap between pure engineering mechanics and strategic business outcomes—covering everything from breaking system bottlenecks to 𝚜̲𝚑̲𝚒̲𝚏̲𝚝̲𝚒̲𝚗̲𝚐̲ ̲𝚕̲𝚎̲𝚏̲𝚝̲ ̲𝚠̲𝚒̲𝚝̲𝚑̲ ̲𝙿̲𝚘̲𝚕̲𝚒̲𝚌̲𝚢̲–̲𝚊̲𝚜̲–̲𝙲̲𝚘̲𝚍̲𝚎̲ ̲𝚊̲𝚗̲𝚍̲ ̲𝚉̲𝚎̲𝚛̲𝚘̲ ̲𝚃̲𝚛̲𝚞̲𝚜̲𝚝̲ ̲𝚜̲𝚝̲𝚊̲𝚗̲𝚍̲𝚊̲𝚛̲𝚍̲𝚜̲. If you are building, leading tech teams, or thinking about distributed systems that must scale sustainably, do yourself a favour and pick up a copy of this book published by Packt 🔗 𝗚𝗿𝗮𝗯 𝘆𝗼𝘂𝗿 𝗰𝗼𝗽𝘆 𝗵𝗲𝗿𝗲: https://lnkd.in/djX8RCte 𝗛𝘂𝗴𝗲 𝗰𝗼𝗻𝗴𝗿𝗮𝘁𝘂𝗹𝗮𝘁𝗶𝗼𝗻𝘀 𝘁𝗼 𝗜𝗺𝗿𝗮𝗻 𝗼𝗻 𝗽𝘂𝘁𝘁𝗶𝗻𝗴 𝘁𝗼𝗴𝗲𝘁𝗵𝗲𝗿 𝘀𝘂𝗰𝗵 𝗮 𝗽𝗿𝗮𝗰𝘁𝗶𝗰𝗮𝗹, 𝗲𝘃𝗶𝗱𝗲𝗻𝗰𝗲-𝗯𝗮𝘀𝗲𝗱 𝗴𝘂𝗶𝗱𝗲! #SystemDesign #CloudArchitecture #TechLeadership #ScaleBySubtraction #StartupFounder #SoftwareEngineering
-
Imran Siddique shared thisWhen your project is new, you are asking people to spend their time before they know whether your work will help them. With AgenTrust, we gave people several places to start: code they could run, a marketplace where their integrations could be found, and research they could examine and challenge. The download history starts at 251 on June 4. By October 4, it reaches 641,058 across all 11 identified published distributions. The graph keeps mirrors, automated installs, repeat downloads and every spike. These are download events, not community members. The more useful story is in the contributions. Dhillon Kannabhiran brought SAGE; Carlos Hernández helped improve its description through review. Val Rukhaylo worked through installation and verification issues to get Agent Replay ready for others to use. Aaron Fulkerson and Rishabh Poddar originated TRACE, with engineering support from OPAQUE. I brought existing relationships too. That context matters when telling a story about starting a new community. Alongside Carlos, Mohammed Zoheb Shaik, Dipika Ranabhat, Louie Lu and Rajnish Tiwari have taken on community maintainer roles. That is another part of the progress worth recognizing. In this edition of Proof, Not Promises, I walk through those examples, the marketplace and the research library, with links you can use. If you are building at a startup, the practical starting point is small: one useful example, a clear review path, and visible credit for the person who improves it. Then give repeat contributors responsibility for maintaining the work. The next progress I want to see is more people reviewing, maintaining and reproducing results independently. #OpenSource #CommunityBuilding #AIAgents Rishabh Poddar Aaron Fulkerson Dhillon Kannabhiran Val Rukhaylo
-
Imran Siddique shared thisAn agent calls a tool. Six months later, someone asks: which policy allowed it, was that the approved agent, and can we still trust the record? I’m speaking at 𝗔𝗚𝗡𝗧𝗖𝗼𝗻 + 𝗠𝗖𝗣𝗖𝗼𝗻 𝗡𝗼𝗿𝘁𝗵 𝗔𝗺𝗲𝗿𝗶𝗰𝗮 in San Jose about how to answer those questions. 𝗚𝗼𝘃𝗲𝗿𝗻𝗲𝗱 𝗠𝗖𝗣: 𝗣𝗼𝗹𝗶𝗰𝘆 𝗘𝗻𝗳𝗼𝗿𝗰𝗲𝗺𝗲𝗻𝘁, 𝗧𝗮𝗺𝗽𝗲𝗿-𝗘𝘃𝗶𝗱𝗲𝗻𝘁 𝗔𝘂𝗱𝗶𝘁 𝗖𝗵𝗮𝗶𝗻𝘀, 𝗮𝗻𝗱 𝗛𝗮𝗿𝗱𝘄𝗮𝗿𝗲-𝗦𝗲𝗮𝗹𝗲𝗱 𝗥𝗲𝗰𝗲𝗶𝗽𝘁𝘀 𝗙𝗿𝗶𝗱𝗮𝘆, 𝗢𝗰𝘁𝗼𝗯𝗲𝗿 𝟮𝟯 · 𝟰:𝟬𝟱–𝟰:𝟯𝟬 𝗣𝗠 𝗣𝗗𝗧 Room 230C · San Jose McEnery Convention Center I’ll walk through: • How Cedar policies govern MCP tool calls before execution. • How audit chains make later changes detectable. • How hardware attestation helps establish what ran. 𝘞𝘦’𝘭𝘭 𝘢𝘭𝘴𝘰 𝘤𝘰𝘷𝘦𝘳 𝘵𝘩𝘦 𝘭𝘪𝘮𝘪𝘵𝘴: what a signed receipt proves, what needs hardware evidence, and what neither can tell you about whether an agent’s answer was correct. If you’re putting agents in front of enterprise data, this session will give you concrete ways to evaluate enforcement and the evidence it produces. Thanks to Agentic AI Foundation, Angie Jones, Demetrios Brinkmann, Ijeoma Onwuka, Mazin Gilbert, MBA, Ph.D., Manik Surtani, Alex Salkever, Kylie Wagar-Dirks and Jake P. for bringing the community together. Conference details: https://lnkd.in/gvEydhxR Use COMMUNITY50 for 50% off registration, courtesy of Demetrios. #AGNTCon #MCPCon #MCP #AgentSecurity #AgenticAI #OpenSource
-
Imran Siddique reacted on thisImran Siddique reacted on thisAn AI agent can stay inside its sandbox and still do the wrong thing. Anthropic documented an example in Claude Cowork. Its egress proxy correctly allowed api.anthropic.com. A malicious file supplied an attacker-controlled API key, and Claude uploaded workspace files to the attacker’s Anthropic account. The request stayed within the permitted network boundary. The effect was still wrong. This is why “sandboxed” cannot become shorthand for “governed.” Agent security requires controls that answer different questions: • Which action should the model propose? • Is that exact effect authorized? • Can the process reach the required resource? • What evidence can another party verify afterward? Decision models, governance policy, sandboxes, and audit evidence each serve a purpose. Hardware attestation can strengthen specific claims about the runtime that produced a record. The difficult engineering work is connecting those controls. The system needs a trustworthy description of what an action will do, a record of what the user intended, and enforcement that binds the approval to what actually executes. My latest article examines those boundaries through a checkout attack and recent work from Microsoft, Anthropic, and TRACE. Your AI Agent’s Sandbox Is Not Its Governance System https://lnkd.in/esEdh4V9 Views are my own.
-
Imran Siddique reacted on thisImran Siddique reacted on thisWe’re pleased to share that Custodian has joined NVIDIA Inception. Unlocking AI’s potential requires confidence in how models are protected, deployed and controlled. Model owners want to bring valuable models to new environments while retaining release authority. Governments and enterprises want those models running on infrastructure they control, under conditions they can independently verify. Custodian brings those interests together through shared custody. Each party holds an independent key share. Both must authorize release to an approved, attested runtime, and continued execution is governed by time-bounded authorization. Either party can independently withdraw from future authorization, without the other’s agreement. Continued execution requires fresh authorization before the current lease expires. Confidential GPU computing is central to where this goes next. Our work so far includes hardware evaluations on H100 and Blackwell configurations, and research into what an NVIDIA attestation chain does and does not establish. Joining NVIDIA Inception is an encouraging milestone, and we look forward to drawing on its technical resources and ecosystem as we continue building and validating Custodian. Thank you to NVIDIA for Startups, and to the contributors and reviewers whose scrutiny has made this work better. Ship the Model. Keep the Keys. Stay Sovereign. https://custodylabs.co #NVIDIA for Startups #NVIDIAInception #SovereignAI #ConfidentialComputing #ModelSecurity #ConfidentialAI
-
Imran Siddique liked thisImran Siddique liked thisWould you trust an AI request handled by a machine you’ve never seen before? 🤔 That’s one of the harder questions we’ll tackle next week! 🚨 Join us Monday, October 12 at 5 PM ET / 2 PM PT for Mesh-LLM & Trusting a Stranger’s Machine, Part 2: Evidence You Can Check. When an AI request runs on someone else’s machine, simply being told “it worked” isn’t enough. So what if you could actually check the evidence? 🔗 Mesh-LLM’s Evidence plugin records signed, hash-chained evidence on each node and sends the signed records to both sides of the exchange. We’ll walk through a demo showing: → Two-sided confirmation → A paid Lightning exchange → How the exchange can be represented alongside a TRACE record → What additional runtime and hardware evidence can tell us → And, importantly, where the limits of this approach are Speakers: Michael Neale is the founder and creator of Mesh-LLM, an open-source framework focused on connecting and orchestrating AI models and agents. His work explores practical approaches to building more flexible, collaborative AI systems and making multi-model workflows easier to develop and deploy. Steven Mih is the founder of Action State, focused on building infrastructure and tools that make AI agents more reliable, observable, and accountable. His work centers on helping organizations understand and manage what agents do as they operate across increasingly complex systems. Imran Siddique is the founder of agentrust-io and works on TRACE, an open specification for portable, verifiable records of AI agent runs. His work focuses on creating a common way to capture and verify what AI agents do, enabling greater transparency, trust, and portability across agentic systems. If AI agents are going to work across machines we don’t control, evidence matters. Register today and do not miss this amazing opportunity to learn from the biggest in the industry: https://lnkd.in/eNk-AmMr
-
Imran Siddique liked thisImran Siddique liked thisSome passions never change, and some journeys are truly inspiring! I've known Imran Siddique since our college days, and one thing I vividly remember about him is his love for books. He always had a book in his hand, apart from the ones already tucked away in his bag! Fast forward to today, and it's amazing to see that same book lover become a published author! And here's something special for me too — after almost 15 years, I'm finally picking up a physical book instead of reading on a screen! And what better way to restart this habit than with a book written by someone you've known since college days, on topics you're genuinely passionate about — System Design and Software Architecture! It's the perfect combination of reconnecting with an old habit, revisiting college memories, and exploring something I truly enjoy. Super excited to get my hands on it and dive in. Looking forward to seeing where this reading journey takes me! Kudos, Imran! It's wonderful to see your passion for reading evolve into a passion for writing and sharing knowledge. Wishing you many more milestones ahead! #SystemDesign #SoftwareArchitecture #SystemArchitecture #SoftwareEngineering #BookReading #ReadingJourney #PublishedAuthor #CollegeMemories #TechBooks #ContinuousLearning
Experience & Education
Licenses & Certifications
Volunteer Experience
-
Member Board of Directors
Upton Crossroads
- 5 years 9 months
Social Services
Imran Siddique was elected to serve on the board of Upton @ Crossroads, a Bellevue association. As a board member, he is tasked with representing the interests of the association and assisting in the development and implementation of its activities and initiatives.
-
Library Advisory Board
City of Bellevue, Washington
- 1 year 6 months
Education
The Bellevue Library Advisory Board makes periodic reports and recommendations to the City Council and city manager related to the scope and quality of library services provided to Bellevue residents by the King County Library System. The board may also advise the KCLS Board of Trustees regarding acquisition, maintenance and disposition of Bellevue library buildings, as well as rules and regulations governing their use. The Library Advisory Board is comprised of seven Bellevue residents…
The Bellevue Library Advisory Board makes periodic reports and recommendations to the City Council and city manager related to the scope and quality of library services provided to Bellevue residents by the King County Library System. The board may also advise the KCLS Board of Trustees regarding acquisition, maintenance and disposition of Bellevue library buildings, as well as rules and regulations governing their use. The Library Advisory Board is comprised of seven Bellevue residents appointed by the mayor with the concurrence of the City Council.
-
Treasurer Board of Directors
STEVENSON ELEMENTARY SCHOOL PTO
- 1 year 10 months
Children
Elected to serve on the board of the Stevenson Elementary School PTA as treasurer.
-
Visiting Professor
upGrad
- 1 year 1 month
Education
As a volunteer, I have contributed to the education and empowerment of others by teaching and designing courses on machine learning, a field that harnesses the power of artificial intelligence to create innovative solutions. I have also created a tech 101 course for program managers, which introduces them to the essential concepts and skills of software engineering, data science, and cloud computing.
-
Co-Chair
Bellevue College Crossroads Preschool
- 1 year 1 month
Education
Co-chair
-
Member Board of Directors
REDMOND PARENT COOPERATIVE PRESCHOOL
- 1 year 2 months
Education
Parent leader for the 3 AM and 4 AM shifts
-
Special Education Teacher
Marathi Shiksha Abhiyan
- 2 years
Children
-
Special Education Teacher
Sarva Shiksha Abhiyan (SSA)
- 1 year
Poverty Alleviation
-
Co President
Bennett Elementary School PTSA
- Present 2 years 2 months
Education
Patents
-
System for monitoring and management of electrical system
Issued WO2023113894A1
A computing system is provided for intelligent monitoring and management of an electrical system. The system receives electrical usage data and electrical production data via a computer network from a plurality of grid agents at a plurality of points across the electrical system; receives a privacy permission from an owner of one of the plurality of grid agents; based on the privacy permission, authorize sharing of a shareable portion of the electrical usage data and/or a shareable portion of…
A computing system is provided for intelligent monitoring and management of an electrical system. The system receives electrical usage data and electrical production data via a computer network from a plurality of grid agents at a plurality of points across the electrical system; receives a privacy permission from an owner of one of the plurality of grid agents; based on the privacy permission, authorize sharing of a shareable portion of the electrical usage data and/or a shareable portion of the electrical production data and restrict sharing of a restricted portion of the electrical usage data and/or a restricted portion of the electrical production data; and executes a prediction model which processes patterns observed in the shareable portion of the electrical usage data and the shareable portion of the electrical production data, and predicts a forecasted aspect of the electrical system; and outputs the forecasted aspect of the electrical system.
Other inventorsSee patent -
Hybrid cloud data control
Issued WO2023101736A1
A remote server computing system is configured to present a user interface with a plurality of deployment configuration options including compute configuration options and data storage configuration options for energy-related data within a hybrid cloud environment. The hybrid cloud environment comprises a cloud-service-managed control plane and a data plane utilizing local compute resources and storage. A data control policy is generated that provides cloud-service-managed governance over at…
A remote server computing system is configured to present a user interface with a plurality of deployment configuration options including compute configuration options and data storage configuration options for energy-related data within a hybrid cloud environment. The hybrid cloud environment comprises a cloud-service-managed control plane and a data plane utilizing local compute resources and storage. A data control policy is generated that provides cloud-service-managed governance over at least a portion of the data plane. The control plane is configured to enforce the data control policy by subjecting at least a portion of the energy-related data to a data transmission restriction or a local storage restriction. The data plane is used to deploy one or more cloud service functions configured to process at least the portion of the energy-related data and output one or more extracted features from at least the portion of the energy-related data to the data plane.
Other inventorsSee patent -
Deploying a function in a hybrid cloud
Issued WO2023101739A1
A remote server computing system is configured to deploy a cloud-service-managed control plane and a cloud service data plane spanning the remote server computing system, a local edge computing device, and a local on-premises computing device connected in a hybrid cloud environment. Energy-related training data is received including a plurality of energy-related training data pairs. A machine learning function is trained using the plurality of training data pairs to predict a classified label…
A remote server computing system is configured to deploy a cloud-service-managed control plane and a cloud service data plane spanning the remote server computing system, a local edge computing device, and a local on-premises computing device connected in a hybrid cloud environment. Energy-related training data is received including a plurality of energy-related training data pairs. A machine learning function is trained using the plurality of training data pairs to predict a classified label for restricted energy-related data that is not accessible to the remote server computing system. The trained machine learning function is deployed to the one or more of the local edge computing device and the local on-premises computing device via the cloud service data plane. The remote server computing system is further configured to receive, via the cloud service data plane, classified output of the trained machine learning function.
Other inventorsSee patent -
Decentralized data platform
Issued WO2022271307A1
Data from data sources may be processed at an edge device. The edge device may generate a local processing result, filter the data, and/or prioritize the data. Accordingly, data is transmitted from the edge device to the data platform, where it may be processed further. For example, a local processing result may be processed at the data platform, such that processing is performed without all of the data source data. In examples, at least a part of such data may remain at an edge device. The…
Data from data sources may be processed at an edge device. The edge device may generate a local processing result, filter the data, and/or prioritize the data. Accordingly, data is transmitted from the edge device to the data platform, where it may be processed further. For example, a local processing result may be processed at the data platform, such that processing is performed without all of the data source data. In examples, at least a part of such data may remain at an edge device. The edge device may maintain a manifest of data stored by the edge device. The data platform may generate an aggregated manifest using manifests from associated edge devices, such that it may be determined where data is stored. As a result, the data platform may redirect requests to an associated edge device when it is determined that requested data is remote from the data platform.
Other inventorsSee patent -
Heterogeneous data platform
Issued WO2022271322A1
In examples, streaming data is received from a data source (e.g., by an edge device associated with a data platform) and is queued and aggregated. Batch data may similarly be received from a data source. The batch data and the aggregated streaming data may be processed to generate metadata accordingly. The data and metadata may be provided to the data platform, where the metadata may be used to update an index and the data may be stored in association with the index. In some instances, the data…
In examples, streaming data is received from a data source (e.g., by an edge device associated with a data platform) and is queued and aggregated. Batch data may similarly be received from a data source. The batch data and the aggregated streaming data may be processed to generate metadata accordingly. The data and metadata may be provided to the data platform, where the metadata may be used to update an index and the data may be stored in association with the index. In some instances, the data may be stored in chunks to facilitate subsequent retrieval of the data. In response to a request for the data, the index may be used to identify relevant data, which may include data associated with batch and/or streaming data formats, thereby enabling the client device to access data having any of a variety of formats from the data platform.
Other inventorsSee patent -
Audit ledger for externality tracking and reporting
Issued WO2022240512A1
The disclosed technology is generally directed to an audit ledger for externality tracking. In one example of the technology, auditing of a virtual sensor is enabled. The virtual sensor is configured to output an externality value based on received telemetry data. The externality values are associated with the at least one particular type of quantifiable technical externality. Signing of the virtual sensor is enabled. Audit information associated with the signing of the virtual sensor is stored…
The disclosed technology is generally directed to an audit ledger for externality tracking. In one example of the technology, auditing of a virtual sensor is enabled. The virtual sensor is configured to output an externality value based on received telemetry data. The externality values are associated with the at least one particular type of quantifiable technical externality. Signing of the virtual sensor is enabled. Audit information associated with the signing of the virtual sensor is stored on a ledger. Periodic aggregated externality values based on the externality values output by the virtual sensor are calculated. The periodic aggregated externality values are stored on a distributed ledger. An audit of the periodic aggregated externality values is enabled. The auditing of the periodic aggregated externality values includes verifying the signing of the virtual sensor by the auditor based on the stored audit information.
Other inventorsSee patent -
Extensible data platform with database domain extensions
Issued US20220292106A1
A computing system that includes one or more server computing devices including one or more processors configured to execute instructions for a domain extensibility module that provides software development tools for building domain extensions for a database platform, and a data ingestion module that provides software development tools for defining a metadata schema for extracting metadata from data files. The one or more processors are configured to receive a set of data from a user computing…
A computing system that includes one or more server computing devices including one or more processors configured to execute instructions for a domain extensibility module that provides software development tools for building domain extensions for a database platform, and a data ingestion module that provides software development tools for defining a metadata schema for extracting metadata from data files. The one or more processors are configured to receive a set of data from a user computing device, define a target metadata schema that includes one or more metadata fields that will be populated during a data ingestion process, define a target domain extension that defines one or more data types for storing the received set of data after performing the data ingestion process, and ingest the received set of data using a metadata extraction pipeline to generate metadata files based on the target metadata schema.
Other inventorsSee patent -
Energy data platform
Issued WO2022182579A1
Examples are disclosed that relate to an energy data platform. One example provides a method comprising receiving a first energy data set having a first data format, and a second energy data set having a second data format, and ingesting the first energy data set and the second energy data set by automatically converting one or more of the first energy data set and the second energy data set into a standard data format. The method further comprises receiving a request from a first application…
Examples are disclosed that relate to an energy data platform. One example provides a method comprising receiving a first energy data set having a first data format, and a second energy data set having a second data format, and ingesting the first energy data set and the second energy data set by automatically converting one or more of the first energy data set and the second energy data set into a standard data format. The method further comprises receiving a request from a first application to provide the first energy data set in the first data format, and in response, providing the first energy data set in the first data format, and receiving a request from a second application to provide the first energy data set in the standard data format, and in response, providing the first energy data set in the standard data format.
Other inventorsSee patent -
Tracking and auditing of ghg emissions
Issued US20220268749A1
One example provides a computing system configured to implement a greenhouse gas (GHG) tracking and auditing platform for tracking GHG emissions associated with a plurality of GHG entities, each GHG entity representing one or more of a GHG source and a GHG sink. The computing system comprises a logic subsystem and a data-holding subsystem comprising computer-readable instructions. The instructions are executable to receive sensor data over time from each sensor of one or more sensors, each…
One example provides a computing system configured to implement a greenhouse gas (GHG) tracking and auditing platform for tracking GHG emissions associated with a plurality of GHG entities, each GHG entity representing one or more of a GHG source and a GHG sink. The computing system comprises a logic subsystem and a data-holding subsystem comprising computer-readable instructions. The instructions are executable to receive sensor data over time from each sensor of one or more sensors, each sensor configured to sense data related to GHG emissions. The instructions are further executable to, for each time interval of a plurality of time intervals, determine a GHG emission amount based at least in part on the sensor data and store GHG emission data. The instructions are further executable to receive a request for a set of GHG emission data and in response to the request, output the set of GHG emission data.
Other inventorsSee patent -
Application program interface for hierarchical data files
Issued US20220269649A1
A server computing device including a processor. The processor may be configured to, via an application program interface (API), receive a hierarchical data file including a plurality of datasets that are hierarchically organized in a plurality of dataset groups. The processor may be further configured to assign respective dataset metadata to the datasets and respective dataset group metadata to the dataset groups. The processor may be further configured to store, in memory, the plurality of…
A server computing device including a processor. The processor may be configured to, via an application program interface (API), receive a hierarchical data file including a plurality of datasets that are hierarchically organized in a plurality of dataset groups. The processor may be further configured to assign respective dataset metadata to the datasets and respective dataset group metadata to the dataset groups. The processor may be further configured to store, in memory, the plurality of datasets, the dataset metadata, and the dataset group metadata. The processor may be further configured to, via the API, receive a dataset query from a client computing device. The processor may be further configured to perform a search over the dataset metadata and/or the dataset group metadata to thereby generate search results. The processor may be further configured to transmit the search results to the client computing device via the API.
Other inventorsSee patent -
Estimating emissions with virtual sensor models
Issued WO2022177694A1
A computing device including a processor is provided. For one or more emissions factors, the processor may be configured to receive respective emissions factor data including sensor data received from one or more sensors and user-specified data received at a user interface. The processor may be further configured to receive a plurality of virtual sensor models. Each virtual sensor model may include one or more respective virtual sensor model parameters. At least in part by applying the virtual…
A computing device including a processor is provided. For one or more emissions factors, the processor may be configured to receive respective emissions factor data including sensor data received from one or more sensors and user-specified data received at a user interface. The processor may be further configured to receive a plurality of virtual sensor models. Each virtual sensor model may include one or more respective virtual sensor model parameters. At least in part by applying the virtual sensor models to the emissions factor data, the processor may be further configured to compute a plurality of virtual sensor readouts. The processor may be further configured to generate a greenhouse gas emissions estimate from the virtual sensor readouts. The processor may be further configured to convey an indication of the greenhouse gas emissions estimate to the user interface for output on one or more output devices.
Other inventorsSee patent -
Skillset scoring and extraction engine
Issued US20210374649A1
Methods and systems are provided that are directed to identifying an individual having a specific skill and that is best suited for performing a particular task. The individual may be identified based on having a highest score for the skill out of a pool of individuals having the same skill. In examples, a score specific to each individual out of the pool of individuals having the same skill may be based on multiple interactions with the skill and may be adjusted based on a recency of when each…
Methods and systems are provided that are directed to identifying an individual having a specific skill and that is best suited for performing a particular task. The individual may be identified based on having a highest score for the skill out of a pool of individuals having the same skill. In examples, a score specific to each individual out of the pool of individuals having the same skill may be based on multiple interactions with the skill and may be adjusted based on a recency of when each interaction occurred and a level of involvement the individual had with the skill. Accordingly, individuals having recent experience with the skill may be scored, and therefore, ranked higher than individuals having experience with the skill that may have occurred further back in time. In examples, the information about the skills and experiences may be utilized to automatically generate the score.
Other inventorsSee patent -
Accessibility controls in distributed data systems
Issued US11128471B2
Methods, systems, apparatuses, and computer program products are provided for controlling data access and/or accessing data in a distributed network. An identifier may be generated for confidential data. A hash generator may generate a first hash of the confidential data, and generate a second hash based on the first hash and the identifier. A data packet set that includes the confidential data, the identifier, and the second hash may be encrypted and transmitted between nodes of the…
Methods, systems, apparatuses, and computer program products are provided for controlling data access and/or accessing data in a distributed network. An identifier may be generated for confidential data. A hash generator may generate a first hash of the confidential data, and generate a second hash based on the first hash and the identifier. A data packet set that includes the confidential data, the identifier, and the second hash may be encrypted and transmitted between nodes of the distributed network. A receiving node may decrypt the packet set using one or more appropriate encryption keys and regenerate the second hash from the confidential data and the identifier. The regenerated hash may be compared against the decrypted hash to authenticate the received confidential data and/or validated against a revocable hash stored in an array of the owning node that is accessible by any receiving node.
Other inventorsSee patent -
Node and cluster management on distributed self-governed ecosystem
Issued US11012506B2
Adding a computing node to a distributed computing system. A method includes executing a binary, for nodes of the distributed computing system, at the computing node. A public node is identified as a result of executing the binary. Various characteristics of the computing node are identified as a result of executing the binary. The various characteristics are provided to the public node. The computing node receives from the public node a registration confirmation packet comprising information…
Adding a computing node to a distributed computing system. A method includes executing a binary, for nodes of the distributed computing system, at the computing node. A public node is identified as a result of executing the binary. Various characteristics of the computing node are identified as a result of executing the binary. The various characteristics are provided to the public node. The computing node receives from the public node a registration confirmation packet comprising information identifying a category from among a plurality of categories of the distributed computing system. The identified category is identified for the computing node based on the identified characteristics of the computing node, and based on the health of the identified category. The computing node is connected to the distributed computing system by the computing node connecting to the identified category.
Other inventorsSee patent
Other 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