À propos
With a proven track record at Google as a Senior Customer Engineer, my mission is to…
Activité
2 k abonnés
-
Tony Carnal a republié ceciTony Carnal a republié ceciBuilding intelligent agents is one challenge, but the real work begins with managing and governing them effectively in production. Google Cloud provides the comprehensive ecosystem developers need for the entire agent lifecycle. From leveraging various agent frameworks (ADK, LangGraph, CrewAI) to integrating with a wide array of tools (Agent2Agent, custom, ecosystem) and powerful models (Gemini), we're giving you the control. Here’s what this means for developers: ✅ Build Your Way: You're not locked in. Use open-source frameworks you already know, like our Agent Development Kit (ADK), LangGraph, CrewAI, LangChain, and more. ✅ Speedup prototype-to-production: Agent Engine is a fully managed runtime that solves the prototype-to-production gap. It handles the tough parts -infrastructure, scaling, security, and even session/memory management, so you can focus on agent logic. ✅ Get Visibility: Understand exactly what your agent is doing. Get deep observability with integrated tools like Cloud Trace and Logging to debug and monitor behaviour. ✅ Connect Anything: Easily plug your agent into a massive ecosystem of tools and data sources, including your own custom APIs. ✅ Maintain Control: Shape agent reasoning with deterministic guardrails and use integrated evaluation tools to ensure your agents perform reliably and safely. It's about empowering you to create reliable, performant agents at scale, without the typical overhead. The attached diagram shows how these pieces fit together ⬇️
-
Tony Carnal a republié ceciTony Carnal a republié ceciThe era of the “toy agent” is over. 🚀 Over the last two weeks, Google Cloud has quietly overhauled the entire agent ecosystem – from the Agent Development Kit (ADK) to the Vertex AI Agent Engine. If you’re an Architect or CTO, this is the moment where agents move from experimental to mission-critical. As someone helping enterprises build GenAI systems every day, these are the 3 shifts that really matter: 1️⃣ Operational maturity 🛠️ The new Vertex AI Agent Engine gives us deep observability: traces, metrics, and a User Simulator to safely test non-deterministic behavior.Pair that with the Go SDK and a one-command adk deploy, and we finally have a clean path from notebook → managed runtime → production. 🔥 The Engine is rolling out to Australia and SE Asia by end of Q4, solving for latency-sensitive workloads. 2️⃣ Self-healing logic 🧠 The updated ADK introduces “self-healing” tool use. Agents can detect when a tool call fails and automatically retry with new parameters or strategies. That’s how you move from “it works in 80% of cases” to something you can trust at scale. 3️⃣ Enterprise governance 🔐 The real game-changer: agents now have first-class IAM identities. Combined with Model Armor, Security Command Center integration, Private VPC, and a generous free tier for Agent Engine, security and cost controls are baked into the platform – not bolted on later. 📚 Want to go deeper? Google just released a 54-page “Introduction to Agents” whitepaper and a free 5-Day AI Agents Intensive with Kaggle . We’re moving fast from “chatbots on documents” to governed, intelligent systems that behave more like microservices. 👉 In your 2026 roadmap, are agents still side projects – or part of your production architecture? #GoogleCloud #VertexAI #AIAgents #GenAI #SoftwareArchitecture #DevOps #CIO #CTO
-
Tony Carnal a republié ceciTony Carnal a republié ceciGoogle has been named a Leader in the 2025 IDC MarketScape for Worldwide GenAI Life-Cycle Foundation Model Software Vendor Assessment! Read why organizations, from innovative startups to the most demanding enterprises, choose Google Cloud for their critical generative AI deployments → https://goo.gle/4oDl5PP
-
Tony Carnal a republié ceciTony Carnal a republié ceciGoogle just released a game-changing 54-page blueprint on AI agents, and it’s not what you think. This isn’t about building smarter chatbots. It’s about creating complete autonomous systems that think, act, and evolve, what Google calls “LMs in a loop with tools to accomplish objectives.” ——— The document breaks down agents into three core components: • The Model (the “Brain”) - for reasoning and decision-making • Tools (the “Hands”) - for taking real-world actions • Orchestration Layer (the “Nervous System”) - managing the Think → Act → Observe loop Google maps out 5 levels of agentic evolution: Level 0: Core reasoning (the model alone) Level 1: Connected problem-solver (model + tools) Level 2: Strategic planner (context engineering + multi-step plans) Level 3: Multi-agent collaboration (teams of specialists) Level 4: Self-evolving systems (agents that create new agents and tools) The Game-Changer: Agent Ops This is where it gets real. Traditional prompt engineering evolves into a full discipline: • Metrics-driven development (business KPIs, not just pass/fail) • LM-as-Judge for quality evaluation • OpenTelemetry traces for debugging agent “thought processes” • Human feedback loops that convert bug reports into permanent test cases Enterprise-Grade Security: Agents introduce a new security principal, not human users, not service accounts, but autonomous actors with their own identity (SPIFFE), policies, and least-privilege permissions. Google’s already thinking ahead: Agent2Agent (A2A) protocols for inter-agent communication, Agent Payment Protocol (AP2) for autonomous transactions, and central governance layers to prevent “agent sprawl.” ——— As an AI leader, I think we’re not just adding AI features anymore. We’re building systems where context window curation becomes an art form, where agents manage other agents, and where the developer shifts from “bricklayer” to “director.” What do you think? 🔃 Hit REPOST and get the full PDF straight to your dms. ___________________________________________ 👋 I am Amit Rawal, an AI practitioner and educator. Outside of work, I’m building SuperchargeLife.ai , a global movement to make AI education accessible and human-centered. ♻️ Repost if you believe AI isn’t about replacing us… It’s about retraining us to think better. Opinions expressed are my own in a personal capacity and do not represent the views, policies, or positions of my employer (currently Google LLC) or its subsidiaries or affiliates.
-
Tony Carnal a republié ceciTony Carnal a republié ceciI don’t think the paper “THE DRAGON HATCHLING” by Pathway, received the attention it deserved, perhaps because it is highly mathematical. The paper introduces BDH, a new architecture that could reshape how language models work by making them more efficient, interpretable, and biologically grounded. In current Transformers, attention works like a classroom where every student talks to every other student before answering a question. It is powerful but extremely inefficient. Each student represents a neuron, a small processing unit, and the connections between them are called synapses. These synapses carry signals from one neuron to another. In a Transformer, every neuron communicates with all others through these synapses, which means the system performs an enormous number of comparisons. This allows the model to capture rich relationships between words but also makes it slow and energy-hungry. BDH introduces linear attention, which changes how this classroom operates. Instead of everyone talking to everyone else, the class appoints a notetaker who keeps a summary of what has already been discussed. When a new word comes in, neurons do not reprocess all previous ones; they simply update the shared notes. This makes context-building continuous and efficient, dramatically cutting down on computation while keeping the meaning intact. The model also borrows from the brain’s balancing act of excitation and inhibition. Imagine that in the classroom, when an important idea appears, a few students become more active and speak louder. That is excitation which strengthens useful signals. But to prevent chaos, some students stay quiet so that the discussion remains clear. That is inhibition which suppresses distractions and irrelevant ideas. BDH applies the same principle mathematically by amplifying strong patterns and dampening weak ones. This keeps reasoning organized and energy use low. When implemented on GPUs, BDH performs as well as much larger Transformer models on translation and text prediction tasks, but it requires fewer parameters and far less energy. Its sparse and focused activity leads to clusters of neurons that specialize in related ideas, giving it an internal structure that is easier to interpret. This work challenges the current “bigger is better” trend in AI. Instead of relying on scale, BDH shows that smarter architectures can achieve similar or even better results with less computation. It addresses major issues in today’s models, such as energy consumption, high cost, and poor interpretability, though questions remain about whether such locally connected systems can handle long or highly abstract reasoning. Still, it feels like an important step forward. The next generation of models may not need to be larger but simply better designed and able to focus, filter, and learn with the same efficiency that makes the human brain remarkable. Paper: https://lnkd.in/exqfG7Ji
-
Tony Carnal a republié ceciTony Carnal a republié ceciToday, we announced Gemini Enterprise - the new front door for AI in the workplace. Gemini Enterprise brings the best of Google AI to every employee through an intuitive chat interface that unifies six core components - - A platform powered by Google’s most advanced Gemini models, providing world-class intelligence for every task; - A no-code workbench that lets any user across the organization analyze information and orchestrate agents to automate processes; - A suite of pre-built Google agents for specialized jobs; - Secure connection to your company’s data wherever it lives – from Google Workspace to Microsoft 365 to business applications like Salesforce and SAP; - A central governance framework so you can visualize, secure and audit all of your agents from one place; - All built on a principle of openness with an ecosystem of over 100,000 partners, ensuring customer choice and fostering innovation. We shared a number of new capabilities across Workspace, Customer Engagement Suite, and new learning resources. Read more on our blog: https://lnkd.in/gyHcrmSU
-
Tony Carnal a republié ceciTony Carnal a republié ceciGartner has positioned Google Cloud as a Leader in the 2025 Magic Quadrant™ for Strategic Cloud Platform Services, ranked highest for Completeness of Vision. This recognition affirms our belief that the future of cloud computing is about providing a comprehensive AI-powered platform for our customers' to solve their most critical challenges. Our vision is centered on four principles: Industrializing AI: Helping companies move AI from one-off projects to a core part of their business process. This includes everything from our custom AI Hypercomputer to the scalable GKE platform. Designing for Efficiency: Eliminating the costs of inefficient infrastructure by co-designing everything - from our custom Axion silicon to our global network - ensuring high performance and predictable costs from the start. Optimizing for enterprise workloads: Our core computing platform brings cost down and improves performance and reliability through deep optimization for top ISV workloads that are critical to our customers. Enabling Global Ambition: We offer tools for digital sovereignty, giving enterprises the control they need to confidently enter new markets and handle complex regulations. Read more on our blog: https://lnkd.in/gQAwW9sbGoogle is a Leader in Gartner® Magic Quadrant for SCPS | Google Cloud BlogGoogle is a Leader in Gartner® Magic Quadrant for SCPS | Google Cloud Blog
-
Tony Carnal a partagé cecigreat opportunity to work with a wonderful team !!Tony Carnal a partagé ceciAre you a Go Backend Engineer with a taste for cloud infras, a music lover, and looking for a new opportunity? Look no more, Deezer is waiting for you. You'll be working on ambitious and innovative new products with a talented team. Being in Pierre-Olivier's team with Florence is your chance to collaborate with Deezer's funniest bassists! https://lnkd.in/dS7MMdTY (If you're a PHP expert, look around on the website, there are other exciting Backend positions to fill as well)
-
Tony Carnal a aimé ceciTony Carnal a aimé ceciHonored on behalf of E-mma organization to be part of the French delegation in India for the AI Summit 🇫🇷🤖🇮🇳👩🏻💻🧑🏻💻 This summit was an important moment to discuss the future of responsible, open, and inclusive artificial intelligence, and the role education must play in preparing the next generation to understand and build these technologies. A sincere thank you to President Emmanuel Macron French Embassy in India / Ambassade de France en Inde Anne Le Hénanff Éléonore Caroit Catherine Vautrin Clara Chappaz Payal S. Kanwar Manisha Khosla Sinha Sanjay Singh Fabien Chareix （夏睿思 / फाबीएं चारिक्स）French Institute India and many others for their support At E-mma, we strongly believe that AI must be an opportunity for everyone, especially for younger generations. Through our educational initiatives such as #CodeAtHome, we aim to give young people the tools to understand, build, and shape the AI of tomorrow. We have reached 90 830 young learners across India, with a strong focus on inclusion: more than 48% of our participants are girls. Our ambition is clearly to empower 1M young people in India to access the skills that will shape the digital world. #AISummit #AI #FranceIndia #AIforGood #Education #indiafrance #franceinindia cc victoire vandeville Aantriksha Chitrey Taru Dahiya Sapna Chadha
-
Tony Carnal a aimé ceciTony Carnal a aimé ceciI’ve been spending couple of hours on a side project: a Google Cloud Knowledge Check service. ☁️ https://lnkd.in/eSHbS3HF The goal is to move beyond static quizzes. This tool uses an AI loop to: 👉 Generate realistic customer scenarios based on specific GCP topics. 👉 Evaluate technical proposals in real-time. 👉 Ground every response in official documentation via the Developer Knowledge MCP Server. (Check it out here https://lnkd.in/eS8BVs53) The best part? I didn’t have to dive deep into the Developer Knowledge API or the MCP server specs—Antigravity handled all the heavy lifting of the integration for me. 🤖 I also took the opportunity to test out Gemini 3.1 Flash-Lite. The speed is fantastic, making the AI loop feel incredibly responsive. ⚡️ If you are a builder or a startup wondering what you could build with Gemini 3.1, I highly recommend checking out Google AI Studio or Antigravity 👇 https://lnkd.in/embvsMhM https://lnkd.in/eDpQXv7H Thanks Romain Vrignaud and Samuel Berthollier for the early feedback ! Anthony Moretti Hadrien Camus Alexandra TRESSOL Guillaume Masson
-
Tony Carnal a aimé ceciTony Carnal a aimé ceci🤩Excited to announce my new role at Google Cloud Consulting organisation as a Pursuit Lead / Engagement Manager! After building a strong foundation of technical skills as a Customer Engineer, I'm excited to start my next chapter by developing my expertise in commercial strategy and deal orchestration. 🚀As a Pursuit Lead, my focus shifts from designing the technical solution to architecting the strategy for customer success. I'll be orchestrating the entire engagement—building the commercial and delivery plan that sets the stage for our delivery teams to succeed and turns a technology vision into a business reality. It's the next step in "compounding" my skills ➡ A recent post from Stephanie Wong on how careers now "fan out, compound, and reset" really resonated with me ➡https://lnkd.in/d7bwPm5z. That's exactly how I see this move. I can't wait to get started! P.S. This photo is a little postcard moment from Dubai, courtesy of #NanoBanana. (And no, I haven’t moved here—just enjoying the visit and the conference!) 🇦🇪✨ #CareerGrowth #GoogleCloud #CareerCompounding
-
Tony Carnal a réagi à ceciTony Carnal a réagi à ceciAfter 6 amazing years at Google, I've decided to take the leap and join Anthropic — and I couldn't be more proud of this decision. Choosing where to work has never felt more meaningful. In a moment when the AI industry is navigating questions about safety, responsibility, and what kind of future we're actually building, I wanted to be somewhere where those questions aren't an afterthought — they're the mission. Anthropic's commitment to AI safety isn't a marketing line. It's baked into how the company thinks, builds, and operates. That matters to me deeply. I'm joining the EMEA team to help bring these capabilities to the enterprises and partners who will shape how AI is adopted across Europe. Big opportunity. Real responsibility. Thrilled to be here. Let's build something worth building. 🤞
-
Tony Carnal a aimé ceciWe finally published it! 🤓 I'm proud to share that Filipe Gracio, PhD and I published the 🚀 Performance Optimization pillar for Google Cloud's AI / ML Architecture Framework. You'll find design guidance on setting realistic goals with clear evaluations metrics to achieve the best model performance and quality leveraging Google Cloud's stack whether you build your own or use our experimentation, training and inference services. Have a good read! 😄Tony Carnal a aimé ceci#GoogleCloud #AI #MachineLearning #CloudArchitecture #GenAI #Strategy #WellArchitected #WellArchitectedFramework 🚀 Scaling AI? Performance is key.🚀 My colleague Benjamin (Haïm) Sadik and I just published the AI/ML Performance Optimization guide for the Google Cloud Architecture Framework. Inside, we cover the 5 core pillars of performance: ✅ Establish Objectives: How to translate business goals into measurable performance and evaluation metrics. ✅ Experiment Frequently: Strategies to run, track, and manage experiments to iterate faster. ✅ Automate Infrastructure: Best practices for building robust training and serving pipelines. ✅ Design for Requirements: How to match your model architecture choices to your specific performance needs. ✅ Link Metrics to Config: Connecting high-level performance metrics directly to low-level design and configuration choices. Read the full guide here: https://lnkd.in/eHCxi4xXAI and ML perspective: Performance optimization | Cloud Architecture Center | Google Cloud DocumentationAI and ML perspective: Performance optimization | Cloud Architecture Center | Google Cloud Documentation
-
Tony Carnal a aimé ceciTony Carnal a aimé ceciWow! I’m approaching my 5-year anniversary as a Customer Engineer at Google 🥂 Reflecting on these five years, it’s been an incredible journey 🤩 🤩 I've met so many brilliant Googlers within Google Cloud sales and throughout the company, 🤝 Helped 100+ EMEA Public Sector customers transform with Google Cloud, 🧑🏫 Trained and empowered users to get the most out of GCP, 🎤 Organized and spoke at countless conferences, both online and offline, 🏅 Dedicated time to student mentorship, startup accelerators, and DEI programs, 📕 Wrote the very first book about #Looker and contributed to the Google Cloud blog. I could go on, but I'll spare you the bragging for now! 😁 𝗪𝗵𝗮𝘁 𝗲𝘅𝗮𝗰𝘁𝗹𝘆 𝗶𝘀 𝗮 𝗖𝘂𝘀𝘁𝗼𝗺𝗲𝗿 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿 (𝗖𝗘) 𝗮𝘁 𝗚𝗼𝗼𝗴𝗹𝗲? Technically, we are the advisors bridging the gap between sales and engineering. We show businesses how Google Cloud solves their specific problems by designing solutions, building prototypes, and delivering demos to prove the tech fits the need (check this link for more info about the role https://lnkd.in/dXyZA6y6). But honestly? This job is so much deeper than I expected. I’ve enjoyed exploring all the unique, creative angles it offers! 🤓 Now, it’s time for a new chapter. I’ll be sharing more about 𝗺𝘆 𝗻𝗲𝘅𝘁 𝗮𝗱𝘃𝗲𝗻𝘁𝘂𝗿𝗲 𝗮𝘁 𝗚𝗼𝗼𝗴𝗹𝗲 very soon! Stay tuned. 🚀 😉
-
Tony Carnal a aimé ceciTony Carnal a aimé ceci📝 𝗥𝗲𝗶𝗻𝗳𝗼𝗿𝗰𝗲𝗺𝗲𝗻𝘁 𝗙𝗶𝗻𝗲-𝗧𝘂𝗻𝗶𝗻𝗴 𝗳𝗼𝗿 𝗟𝗟𝗠𝘀 𝗮𝗻𝗱 𝗔𝗴𝗲𝗻𝘁𝘀: 𝗣𝗮𝗿𝘁 𝟮 - 𝗪𝗵𝗲𝗻 𝗮 𝗦𝗺𝗮𝗹𝗹 𝗣𝗹𝗮𝗻𝗻𝗲𝗿 𝗔𝗴𝗲𝗻𝘁 𝗢𝘂𝘁𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝘀 𝗮𝗻 𝗔𝗴𝗲𝗻𝘁 𝗕𝘂𝗶𝗹𝘁 𝘄𝗶𝘁𝗵 𝗟𝗮𝗿𝗴𝗲𝗿 𝗠𝗼𝗱𝗲𝗹𝘀 👉 Article: urlr.me/PrJQeD 👉 Repo: urlr.me/EKT7nP I just published Part 2 of my guide on Reinforcement Fine-Tuning for LLMs and Agents, and this time it comes with a full repo to try it yourself. 🎯 𝗧𝗵𝗲 𝗽𝗿𝗼𝗯𝗹𝗲𝗺 In Part 1, I covered the fundamentals of RFT. Part 2 tackles a concrete question: can you train a small reasoning model to make better orchestration decisions than a bigger, more expensive one? The use case is a retail customer service system built on a modified version of the tau-bench dataset. Customers come in with all kinds of requests like returns, exchanges, order modifications and a planner agent has to decide which specialist agents are needed to handle each one. The planner doesn't execute anything itself, it's connected to a ReAct agent that does the work. 🔬 𝗧𝗵𝗲 𝗮𝗽𝗽𝗿𝗼𝗮𝗰𝗵 I wrote a Python grader using the F2 metric that is heavily recall-weighted, because missing a required tool is worse than suggesting an extra one. Then I trained o4-mini with low reasoning effort through Azure OpenAI's RFT pipeline. The grader design wasn't straightforward though. The article goes into the iterations, the reward hacking I ran into with a bad grader, and why binary scoring kills the learning signal. Honestly that's where most of the work went. 📊 𝗧𝗵𝗲 𝗿𝗲𝘀𝘂𝗹𝘁𝘀 The fine-tuned o4-mini outperformed every configuration I tested, including gpt-5.2 with high reasoning effort. F2 score of 0.95, recall at 0.97. And it does this with 54% fewer reasoning tokens than vanilla o4-mini and 2.5× lower latency. The model internalized the task well enough that it reasons less and decides better. 🛠️ 𝗧𝗿𝘆 𝗶𝘁 𝘆𝗼𝘂𝗿𝘀𝗲𝗹𝗳 The repo has the source code, training data, grader, and Jupyter notebooks that walk you through the full pipeline: training, evaluation, and running the multi-agent system. You can reproduce the results or adapt it to your own use case.
Expérience et formation
-
Google
Voir toute l’expérience de Tony
Découvrez son poste, son ancienneté et plus encore.
ou
Licences et certifications
-
-
Cambridge CAE (B)
Cambridge ESOL
Délivrance le -
-
-
-
Professional Cloud Security Engineer
Google Cloud
-
-
Professional Cloud Network Engineer
Google Cloud
-
Cloud Digital Leader
Google Cloud
-
Professional Data Engineer
Google Cloud
Langues
-
French
Bilingue ou langue natale
-
English
Capacité professionnelle complète
Recommandations reçues
-
Membre LinkedIn
8 personnes ont recommandé Tony
Inscrivez-vous pour y accéder