Join the 2-Day Free AI Upskilling Sprint by Outskill which comes with 16 hours of intensive training on AI frameworks, tools and tactics that will make you an AI expert. Originally priced at $499, but the first 100 of you get in for completely FREE! Claim your spot now for $0! 🎁

Inside the AI Bootcamp, you will learn:

You will be learning from mentors from the top industries across the globe like Microsoft, Google, META, Amazon, etc. 🎁 You will also unlock $3,000+ in AI bonuses: 💬 Slack community access, 🧰 top AI tools, and ⚙️ ready-to-use workflows — all free when you attend!

Sponsored

Subscribe | Submit a tip | Advertise with us

Welcome to DataPro #137, where the frontier of applied AI meets hands-on innovation.

This week, we’re diving into the tools and breakthroughs reshaping how developers build, deploy, and evaluate intelligent systems, from hyper-efficient LLM fine-tuning to infrastructure-grade agent orchestration. Whether you're launching AI apps on Cloud Run, translating PyTorch to GPU kernels, or exploring agentic architectures, the latest edition has you covered.

What’s New and Noteworthy?

Stay curious, build fast, and experiment responsibly.

Cheers,

Merlyn Shelley

Growth Lead, Packt

Get a head start on our upcoming release, Mathematics of Machine Learning by Tivadar Danka, with this free downloadable primer.

🔍 Inside:

📩 Enter your email to get Essential Math for Machine Learning delivered to your inbox within 24 hours.

Wysh Life Benefit lets banks offer free life insurance through savings accounts, no forms, no opt-ins, no fees. Coverage grows with deposits and is paid directly to the account if the unexpected happens.

Proven impact:

Smart, simple, and built for modern banking.

Sponsored

⏩ Optimizing Multi-Objective Problems with Desirability Functions: This blog explores how desirability functions can simplify complex decisions involving multiple, often conflicting goals. By transforming different metrics into a common scale, it becomes easier to find balanced solutions that meet all requirements. Through a relatable bread-baking example and practical Python code, the post offers a flexible approach to optimizing real-world scenarios, whether in product development, resource planning, or everyday problem-solving.

⏩ Agentic AI 102: Guardrails and Agent Evaluation: Build safer, smarter AI agents by integrating guardrails, evaluation metrics, and real-time monitoring. As AI systems become more autonomous, ensuring they behave reliably and stay on-topic is critical. This blog walks through how to restrict unwanted outputs with Guardrails AI, assess agent quality using DeepEval, and track agent performance via Agno’s monitoring dashboard. With clear examples and code, it’s a practical guide to creating AI agents that are not only capable, but also controlled, transparent, and trustworthy.

⏩ Understanding Random Forest using Python (scikit-learn): Learn how to build accurate, interpretable machine learning models using Random Forests in Python with scikit-learn. This blog breaks down how ensemble methods improve prediction, how to train and tune a Random Forest, and how to assess feature importance. With hands-on code, visualizations, and real-world data, this guide helps you confidently apply Random Forests to both classification and regression problems.

⏩ Google AI Releases MedGemma: An Open Suite of Models Trained for Performance on Medical Text and Image Comprehension. Explore Google’s new MedGemma models to power the next generation of medical AI applications. Unveiled at Google I/O 2025, MedGemma combines advanced text and image comprehension to support diagnostics, clinical reasoning, and medical image interpretation. This blog outlines the model’s capabilities, open-access deployment via Hugging Face and Vertex AI, and how developers can fine-tune it for real-world healthcare use cases.

JULY 16–18 | LIVE (VIRTUAL)

20+ ML Experts | 25+ Sessions | 3 Days of Practical Machine Learning and 35% OFF

Use CodeEARLY35at checkout

Learn Live fromSebastian Raschka,Luca Massaron,Thomas Nield, and many more.

35% OFF ends soon – this is the lowest price you’ll ever see.

⏩ AI Edge Portal brings on-device ML testing at scale: Test and benchmark ML models across real mobile devices at scale with Google AI Edge Portal, now in private preview. This new tool helps developers assess model performance on 100+ Android device models, without the need for a physical lab. The blog explains how the portal simplifies testing cycles, detects hardware-specific issues early, and offers rich performance insights via an interactive dashboard. It’s a game-changer for developers deploying ML at the edge, offering speed, scalability, and actionable data to optimize on-device models.

⏩ Build a domain‐aware data preprocessing pipeline: A multi‐agent collaboration approach. Automate and scale unstructured data processing with a domain-aware, multi-agent pipeline. This blog walks through a robust solution for ingesting and transforming diverse formats like PDFs, transcripts, images, and videos using specialized agents for classification, conversion, and metadata extraction. Built with Amazon Bedrock Agents, the architecture supports modular scalability, human-in-the-loop validation, and continuous improvement, ideal for industries like insurance where accurate metadata fuels analytics, fraud detection, and customer insights

⏩ How public safety agencies can meet AI data security requirements? Safeguard sensitive public safety data while using generative AI with AWS’s secure, compliant infrastructure. This blog outlines how AWS enables agencies to deploy AI responsibly by ensuring full control over data, encrypted communications, and network isolation through technologies like the Nitro System and Amazon Bedrock. It highlights key considerations, including CJIS compliance, access control, and private connectivity, that public safety teams must evaluate when choosing an AI provider. With AWS, agencies can harness AI’s benefits without compromising on privacy, ethics, or security.

⏩ Step-by-Step Guide to Create an AI agent with Google ADK: Build a custom financial analysis agent using Google’s open-source Agent Development Kit (ADK). This blog offers a step-by-step guide to creating an AI agent equipped with tools to fetch real-time company overviews and earnings data using Alpha Vantage APIs. You’ll learn how to structure your project, configure API keys, define tools, and run the agent locally through a browser-based interface. With clear examples and modular code, this walkthrough makes it easy to get started with multi-agent systems using ADK.

⏩ Magentic-UI, an experimental human-centered web agent: Collaborate with AI to complete complex web tasks using Magentic-UI, a human-centered, open-source agent system. Built by Microsoft Research, Magentic-UI blends transparency with control, enabling real-time task execution in your browser with features like co-planning, co-tasking, action guards, and plan learning. Unlike fully autonomous agents, it invites users into the process, offering oversight, adaptability, and safety as core design principles. This blog explores its capabilities, architecture, and how it supports researchers and developers in building more intuitive and responsible AI interactions on the web.

⏩ Predicting and explaining AI model performance: A new approach to evaluation. Predict and explain AI model performance before deployment using ADeLe, a new evaluation framework from Microsoft Research. This blog introduces a novel ability-based approach that rates the cognitive and knowledge demands of tasks, matches them to model capabilities, and forecasts success or failure with high accuracy. By generating detailed ability profiles across 18 scales, ADeLe not only reveals model strengths and weaknesses but also explains why performance varies, offering a powerful tool for developers, researchers, and policymakers seeking more transparent, reliable AI evaluation.

⏩ Introducing Codex: Delegate coding tasks to Codex, a cloud-based AI software engineering agent now available in ChatGPT. Powered by codex‑1 and trained on real-world coding challenges, Codex can write features, fix bugs, propose pull requests, and answer codebase questions, all in parallel cloud environments tailored to your repo. This blog introduces how Codex works, its built-in safeguards, use cases from companies like Cisco and Superhuman, and how developers can begin experimenting today. With task tracking, test logs, and customizable guidance files, Codex brings scalable, asynchronous collaboration to modern software workflows.

⏩ AI Studio to Cloud Run and Cloud Run MCP server: Deploy AI apps in seconds with Cloud Run’s new integration with Google AI Studio and MCP-compatible agents. This blog introduces streamlined tools that let you launch apps with one click from AI Studio, scale Gemma 3 models instantly on Cloud Run with GPU support, and enable AI agents to deploy via the new Cloud Run MCP server. Whether you're prototyping in Gemini, coding in VS Code, or building with agent SDKs, these updates make it easier than ever to build, deploy, and scale AI-powered applications with secure, cost-effective infrastructure.

⏩ Expanding Gemini 2.5 Flash and Pro capabilities: Build smarter, more secure AI solutions with Gemini 2.5 Flash and Pro on Vertex AI. Unveiled at Google I/O, these advanced models introduce features like thought summaries for transparency, Deep Think mode for complex reasoning, and enhanced defenses against prompt injection, making them ideal for enterprise use. Gemini 2.5 is already helping companies like Geotab, Box, and LiveRamp reduce costs, boost accuracy, and scale insights from unstructured data. With generous free credits and seamless integration on Vertex AI, it's now easier than ever to deploy powerful AI across your business.

⏩ A Step-by-Step Coding Guide to Efficiently Fine-Tune Qwen3-14B Using Unsloth AI on Google Colab with Mixed Datasets and LoRA Optimization: Fine-tune large language models like Qwen3-14B efficiently on Google Colab using Unsloth AI. This tutorial walks through a low-resource method for customizing state-of-the-art models using 4-bit quantization and LoRA optimization. With tools like FastLanguageModel, SFTTrainer, and mixed datasets for reasoning and instruction tasks, you can train powerful AI models on consumer-grade GPUs. It’s a practical guide for developers aiming to build custom assistants or domain-specific models without the heavy cost or complexity typically required for LLM fine-tuning.

⏩ Meta Introduces KernelLLM: An 8B LLM that Translates PyTorch Modules into Efficient Triton GPU Kernels: Accelerate GPU programming with KernelLLM, Meta’s new 8B model that translates PyTorch modules into Triton kernels. Fine-tuned from Llama 3.1 and trained on 25K code pairs, KernelLLM simplifies GPU development by automating kernel generation. It outperforms much larger models like GPT‑4o in benchmark tests (Pass@1: 20.2), making GPU acceleration more accessible for developers. Ideal for optimizing deep learning workloads without writing low-level code, KernelLLM represents a major step toward democratizing efficient GPU programming.

⏩ Sampling Without Data is Now Scalable: Meta AI Releases Adjoint Sampling for Reward-Driven Generative Modeling. Train generative models without data, Meta AI’s Adjoint Sampling makes it possible. Tackling the challenge of data-scarce environments, this new algorithm replaces labeled datasets with scalar reward signals, like energy scores from molecular simulations. By modeling sample evolution through stochastic differential equations and optimizing via a novel Reciprocal Adjoint Matching loss, Adjoint Sampling produces high-quality outputs with minimal computation. It scales effectively, respects molecular symmetries, and outperforms traditional methods in energy efficiency and conformer diversity. This breakthrough opens the door for powerful generative modeling in physics, chemistry, and other domains where direct data is hard to come by.

⏩ A Step-by-Step Guide to Deploy a Fully Integrated Firecrawl-Powered MCP Server on Claude Desktop with Smithery and VeryaX: Deploy Firecrawl with Claude Desktop in minutes using MCP, Smithery, and VeryaX. This step-by-step setup connects Firecrawl—an intelligent document-crawling agent, directly to Claude via a fully managed MCP server. Using Smithery’s declarative config and VeryaX’s orchestrated runtime, developers can register APIs, wire up MCP endpoints, and run Firecrawl in Claude’s interface. With just a few commands and API keys, the system integrates real-time scraping capabilities into Claude, enabling contextual AI workflows without custom infrastructure. It’s a scalable, extensible setup for building smarter, data-rich desktop agents.