How to use SageMaker and Comet for ML environments

In this post, we showed how to use SageMaker and Comet together to spin up fully managed ML environments with reproducibility and experiment tracking capabilities.

In this post, we show how to deploy gpt-oss-20b model to SageMaker managed endpoints and demonstrate a practical stock analyzer agent assistant example with LangGraph, a powerful graph-based framework that handles state management, coordinated...

In this post, we show how to deploy gpt-oss-20b model to SageMaker managed endpoints and demonstrate a practical stock analyzer agent assistant example with LangGraph, a powerful graph-based framework that handles state management, coordinated...

Enhance Your Machine Learning Workflow with DeepSeek in SageMaker Studio   https://lnkd.in/e3JT9rJP   Machine learning teams and data scientists working in AWS environments can supercharge their productivity by integrating DeepSeek with SageMaker Studio. This powerful combination transforms how you build, train, and deploy ML models by streamlining complex workflows and automating repetitive tasks.

New version of MLflow - 3.4.0 - has been realeased with lots of new goodies: New Metrics, MCP, Judges & More Key Highlights: • 📊 OpenTelemetry Metrics Export: span‑level stats in OT metrics • 🤖 MCP Server Integration: AI assistants now talk to MLflow • 🧑⚖️ Custom Judges API: Build domain‑specific LLM evaluators • 📈 Correlations Backend: Store & compute metric correlations via NPMI • 🗂️ Evaluation Datasets: Track eval data in experiments • 🔗 Databricks Backend: MLflow server can use Databricks storage • 🤖 Claude Autologging: Auto‑trace Claude AI calls • 🌊 Strands Agent Tracing: Full agent workflow instrumentation • 🧪 Experiment Types in UI: Separate classic ML and GenAI experiments MLflow’s 3.4.0 brings a suite of features that tighten the feedback loop between data scientists and engineers. The OpenTelemetry metrics export gives you end‑to‑end visibility into each span’s performance, while the new MCP server lets LLM‑based assistants query and record runs directly in the tracking store. Custom judges let you author domain‑specific LLM evaluators, and the correlations backend now stores NPMI scores so you can compare metrics across experiments. Versioned evaluation datasets keep all your test data tied to the run that produced it, ensuring reproducibility. The Databricks backend unlocks native Databricks integration for the MLflow server, and auto‑logging for Claude interactions means conversations are captured without manual instrumentation. Strands Agent tracing adds end‑to‑end monitoring for autonomous workflows, and the UI now supports experiment types to keep classic ML/DL work separate from GenAI projects. Full release notes - https://lnkd.in/dtuwVzPk #MLflow #OpenTelemetry #Databricks

I’ve just published a new article on Medium: Building AI Agents with MCP and Amazon Bedrock: From Basics to Real-World Apps Over the last few weeks, I explored how the Model Context Protocol (MCP), Amazon Bedrock, and the Strands agent framework can be combined to build modular and scalable AI systems. To validate the concepts, I developed a series of POCs and documented them in detail. The article walks through: ✅  A baseline agent with Amazon Bedrock (Claude) ✅  Enhancing agents with local tools ✅  Connecting to pre-built MCP servers (AWS Docs, Pricing) ✅  Orchestrating multiple servers to answer complex questions ✅  Building a custom MCP server (Calculator) ✅  A Streamlit demo integrating Kite MCP for portfolio analysis The full GitHub repository with all examples is linked inside the article. I see MCP as an important step towards building enterprise-ready, agentic AI architectures. Would be glad to hear your thoughts. Read the article 👇 #MCP #AmazonBedrock #AIagents #AgenticAI #GenerativeAI #AIArchitecture

Spent about 5 hours on OpenAI’s Agent Builder. Built my own version of a GPTs-style “Butcher” agent, but with a custom interface. As a supposed n8n/Make killer — that’s a joke. 😂 It’s all designed for front-end integration. I hacked together a simple front-end with chat; had to wrestle with it because the front-end turned out to be sensitive to the agent’s architecture (which it really shouldn’t be). At first, there’s a lot of work just getting the front-end running (thankfully Claude Code helps), but then you hit the usual problems with the agent itself. Basically, it’s kind of like n8n/Make — except instead of integrations, you use MCP — and, of course, there are zero services compared to real low-code platforms. It’s not much of a product yet. Especially since Make already has built-in code blocks (and n8n/Zapier have had them for ages). Looks more like they’re testing community reactions. Overall — not great. Complex business logic is still a thousand times easier to build in n8n/Make/Zapier, where you can bridge to any front-end via API/Webhook or any trigger — no headache with complicated protocols. For now, building production-level business logic on this seems dubious. I thought the AI would generate the workflows itself — but nope, still manual. 🤣

Machine Learning Model Deployment: From Jupyter Notebook to Production Your ML model works perfectly in a Jupyter notebook. Then production hits, and everything falls apart. After deploying dozens of ML models from prototype to production, here's the reality: the notebook-to-production gap kills more projects than bad algorithms ever will. The 5-step deployment framework that actually works: 1. Containerize from Day One Don't wait until deployment to think about Docker. Build your model training and inference pipelines in containers from the start. This eliminates the classic "works on my machine" nightmare. 2. Separate Training from Inference Your training pipeline and serving API are completely different beasts. Design them independently with clear interfaces. Training can be slow and resource-heavy; inference must be fast and lightweight. 3. Version Everything Model versions, data versions, code versions—track them all. When your model behaves unexpectedly in production, you need to trace back exactly which components created it. 4. Build Monitoring Before Deployment Track model performance metrics, input data drift, and prediction distributions in real-time. Production models degrade silently—you won't know unless you're watching. 5. Start with Simple Infrastructure Skip the fancy MLOps platforms initially. A REST API with proper logging, a model registry, and basic CI/CD will serve 90% of use cases. Scale complexity only when needed. Last year, a client's recommendation model sat in a notebook for 8 weeks because "deployment was too complex." We got it live in 3 weeks using this framework. The hard truth? A decent model in production beats a perfect model in a notebook every single time. What's been your biggest challenge moving ML models to production? #MachineLearning #MLOps #DataScience #SoftwareEngineering

Machine learning operations (MLOps) is the combination of people, processes, and technology to productionize ML use cases efficiently. To achieve this, enterprise customers must develop MLOps platforms to support reproducibility, robustness, and...

Machine learning operations (MLOps) is the combination of people, processes, and technology to productionize ML use cases efficiently. To achieve this, enterprise customers must develop MLOps platforms to support reproducibility, robustness, and...