Sandro Volpicella’s Post

Simplify getting agents into production OpenAI and AWS partnered -- and have OpenAI-powered agents running on AWS Bedrock The burden is on dev teams to figure out how state is stored. The Stateful Runtime Environment is designed to solve that. 🔗 https://lnkd.in/e6BYjG-v

Missed our live masterclass on ML Deployment on AWS – From Notebook to Production? No worries — the recording is now available. In this session, we covered how to take a machine learning model from a notebook environment to real production using AWS services like EC2, SageMaker, and Lambda, along with a live deployment demo. If you are learning AI / ML / MLOps, this is one of the most important practical skills to understand. Watch the full replay here: https://lnkd.in/gBUQfwN3 #MLOps #AWS #DataScience #MachineLearning #Webinar #CloudComputing #agilefever #masterclass AgileFever

Built my first AI-powered Serverless Application on AWS! I recently built an AI Blog Generator using AWS cloud services. 🔹 Tech Stack: • AWS Lambda • Amazon Bedrock (DeepSeek Model) • API Gateway • Amazon S3 • Python (boto3) How it works: A user sends a blog topic through an API request. The Lambda function calls Amazon Bedrock to generate a blog, and the output is automatically stored in an S3 bucket. This project helped me understand: • Serverless architecture • AI model integration in cloud apps • API-based workflows GitHub Repo: https://lnkd.in/dbXZGjrC Thanks to AWS Free credits and Krish Naik for his amazing tutorial about the Gen AI applications on AWS, Excited to keep building more AI + Cloud projects! #AWS #Serverless #ArtificialIntelligence #CloudComputing #MachineLearning #Python #GenAI

Running unknown code is risky. It can: * break your system * leak secrets * consume resources So I built something to fix this: 👉 SandboxShift It runs code in an isolated sandbox instead of directly on your system. And if your machine doesn’t have enough resources, it automatically runs in your own AWS — not ours. (See demo below 👇) Still in beta — setup is a bit involved (Podman, Terraform, AWS), but I wanted to share early and get feedback https://lnkd.in/gMvdWAa6 Would love to know: *Is this a real problem for you? *Would you use something like this? #DevOps #AI #AWS #OpenSource

Last week, I went hands-on with Amazon Bedrock AgentCore using Strands Agents + Nova Pro. I got to explore what it actually takes to run production-grade AI agents with the BeSA Agentic AI on AWS program. Some things I put into practice: • Deploying agents with AgentCore Runtime for scalable, serverless execution • Enabling agents to run Python and automate web tasks using Code Interpreter and Browser tools • Securing external API access with AgentCore Identity instead of hardcoded credentials • Turning APIs into agent tools with AgentCore Gateway + MCP servers • Adding memory (the coolest part), observability, and tracing to monitor agent behavior in production Building an agent is actually the easy part...but curating an infrastructure that lets agents run securely, remember context, use tools, and scale is the real architecture challenge. 🤖☁️

I have just completed week three of the #BeSA program - Become a solutions architect - Designing with Strands agents What did I learn 🧩 Strands Agents is AWS's open-source SDK for building production-ready AI agents in a few lines of code. The formula is simple: Model + Prompt + Tools. No complex workflow scaffolding. The LLM reasons, selects tools, and drives execution autonomously — already powering Amazon Q Developer and AWS Glue in production. 🔧 Turn any Python function into an agent tool with a single decorator. The @tool decorator is all it takes. The function's docstring becomes the model's instruction — it tells the LLM what the tool does and when to use it. Clean, minimal, and immediately practical. 🔁 The agentic loop is the core engine. Strands runs a continuous Think → Act → Observe → Reflect cycle until the task resolves. No manual orchestration. The model drives each iteration, deciding what to do next based on tool results — including replanning mid-execution when something unexpected comes back. 🤝 Multi-agent patterns are first-class primitives in Strands 1.0. Agents-as-Tools enables hierarchical delegation — an orchestrator agent calls specialist agents the same way it calls any other tool. Swarm and Graph patterns support parallel coordination and complex pipelines without bespoke glue code. 🌐 Model-agnostic by design. Swap between Amazon Bedrock, Anthropic, #OpenAI, #Meta Llama, #Ollama, or #LiteLLM in a single line. No rewriting agents. No framework lock-in. The architecture stays the same regardless of which model sits underneath. 🔭 Observability is built in, not bolted on. #OpenTelemetry support gives you full visibility into each reasoning step, tool call, and decision path. In production, you need to know not just what the agent did — but why. Strands makes that auditable by default. 🛠️ Built a multi-tool Strands agent from scratch in the hands-on workshop — wiring custom tools, running the agentic loop locally, and connecting to MCP servers. Watching the model reason through a multi-step task in real time — selecting tools, evaluating results, replanning — is where the architecture stops being abstract. #StrandsAgents #AgenticAI #AWSBedrock #SolutionsArchitect #OpenSource #CloudArchitecture

Part 2 of my AWS Machine Learning series is out! This time, I focused on the practical side of ML in AWS, how these services are actually implemented in real-world applications. I walk through use cases with Rekognition, Comprehend, Polly, Transcribe, and Translate, and how they integrate into simple cloud architectures using S3 and Lambda. What stood out to me is how fast you can go from idea to implementation. AWS removes a lot of the complexity, allowing developers to focus on solving problems rather than building models from scratch. For certification prep, this is key: understanding how services are used in context. Grateful to Dr. Victor Govindaswamy for his guidance and for making this learning process much more approachable. Watch here https://lnkd.in/gYyb-Xar #AWS #CloudComputing #MachineLearning #Serverless #AWSLambda #S3 #AWSCertification #CloudPractitioner #SolutionsArchitect #AI #TechLearning

Strands Agents | BeSA Workshop – Week 3 ✅ Just completed Week 3 of the BeSA workshop on Strands Agents, and I learned many new things. Strands Agents is an open-source framework for building AI agents that can interact with AWS services and perform complex tasks. Using Amazon SageMaker Studio (JupyterLab), I Created an agent, Added tools, Built a task-specific agent #strandsagents #strandsagentsTools #AWS #AIagents At a high level, the process starts with defining tools. There are multiple ways to define tools using the Strands Agents SDK. One common approach is to use the @tool decorator on a function and provide proper documentation for it. Next, we create the agent. This involves defining a system prompt that specifies what the agent should and should not do. We then configure the agent’s underlying LLM model and provide it with both built-in and custom tools. Strands Agents natively integrate with Amazon Bedrock models. If no model is explicitly defined, the agent falls back to a default LLM. You can also enable thinking mode in the code, allowing the model to manage its own chain-of-thought reasoning. Once everything is configured, we define and invoke the agent. There is still so much to explore, including MCP integration, Langfuse observability, and multi-agent swarms. A big thank you to the awesome AWS team Prasad Rao , Ashish Prajapati , Dr Anil Kumar Giri for designing this program and giving me the opportunity to practice hands-on with AWS infrastructure.

One thing I’ve noticed working on ML systems: the model itself isnt the hardest part. It’s the infrastructure around it ; how data flows in, how quality is checked, how the model gets to production and stays there. Thats where most of the real work is. Here’s the architecture i put together for a ML system: → Data pulled from MangoDB through a 4-stage pipeline → Containerized with Docker, image stored in ECR, compute on EC2 → Model experiments are versioned using MLFlow and Dagshub → Models and artifacts stored in S3 bucket → Model inference is served through an API endpoint using FastAPI → CI/CD is handled through Github Actions The whole thing is designed so any stage can be changed or scaled without touching the rest. #mlsystem #mlops #cloud #aws #docker #fastapi #dagshub #github #githubactions #mongodb #mlarchitecture

🚀 🚀 Automated the one-time AWS account infrastructure setup. Created Terraform Script to deploy (NLB, ALBs, VPC Links, API Gateway, custom domains, domain name associations, ACM certificates, and Route53 records) all provisioned through a single CI/CD pipeline with GitHub Actions. What used to be hours of manual console work across multiple AWS services is now a repeatable, version-controlled git push. Infrastructure as Code isn't optional anymore. It's how you eliminate Data and Infrastructure drift, reduce human error, and onboard new accounts in minutes instead of days. #Terraform #AWS #InfrastructureAsCode #Automation 🔥 🔥