Still don’t use AI to automate your work & make big $$? You’re way behind in the AI race. But worry not: Join the World’s First 16-Hour LIVE AI Upskilling Sprint for professionals, founders, consultants & business owners like you.

Date: 2nd-3rd-4th of May, 11 AM - 7 PM EST.

Rated 4.9/10 by global learners – this will truly make you an AI Generalist that can build, solve & work on anything with AI.

In just 16 hours & 5 sessions, you will:

✅ Learn the basics of LLMs and how they work.
✅ Master prompt engineering for precise AI outputs.
✅ Build custom GPT bots and AI agents that save you 20+ hours weekly.
✅ Create high-quality images and videos for content, marketing, and branding.
✅ Automate tasks and turn your AI skills into a profitable career or business.

All by global experts from companies like Amazon, Microsoft, SamurAI and more. And it’s ALL. FOR. FREE. 🤯 🚀

Join now and get $3000+ in additional bonuses: AI community access ($1999), AI Tool Stack ($299), and Workflow Templates ($999)—all unlocked when you sign up and attend! 🎁

Sponsored

Subscribe | Submit a tip | Advertise with us

Welcome to the 135th edition of DataPro ~ where data meets intelligence, and curiosity powers discovery.

This week, we’re exploring a major leap: data science stepping into the multimodal age. From Google’s BigQuery DataFrames 2.0 transforming Python workflows with built-in support for text, audio, and image data, to open-source audio models and memory-persistent AI agents, the boundary between structured and unstructured analysis is officially dissolving.

🔍 Check out BigQuery DataFrames 2.0 to see how it’s bridging SQL with generative AI, embeddings, and natural language prompts, all while keeping the simplicity of Python dataframes.

But that’s not all. In this issue, we also cover:

🧠 Qwen3’s new switchable thinking modes for better reasoning and conversation

🎧 Kimi-Audio’s low-latency, all-in-one model for audio Q&A, ASR, and more

📚 Anthropic’s dataset on real-world AI value expression

🧰 Building your own agent memory using Claude and knowledge graphs

🧪 Customizing Amazon Nova for accurate, tool-aware AI agents

🧵 Topic model evaluation for business intelligence with FASTopic vs BERTopic

🔄 Evaluating Bedrock Agents using Ragas and LLM-as-a-judge

Plus, hands-on implementations like:

PraisonAI’s fully autonomous data analysis workflows using Gemini

DORA's latest report reveals how generative AI is already reshaping software development

Whether you're scaling up pipelines, testing new models, or integrating agents into real workflows, this edition has ideas worth saving.

We’re also excited to spotlight oursponsorsthis week:Whiteswan Identity Security, delivering zero-trust PAM protection for human and non-human identities across on-prem and cloud environments through a single console.

Meanwhile,HubSpot’s AI-powered ecosystemis expanding rapidly, with a projected global opportunity of $10.2 billion by 2028. To fuel that growth, HubSpot is opening its platform further, introducing an expanded set of APIs, customizable app UIs, and tools that support a more unified data strategy.

Dive in now!

Cheers,

Merlyn Shelley

Growth Lead, Packt

⭕ Qwen/Qwen3-235B-A22B · The newest Qwen3 model brings a major boost in reasoning, coding, multilingual support, and tool use. It smartly switches between deep thinking and fast dialogue, enabling better performance across tasks from chat to agents while supporting over 100 languages and dynamic user control.

⭕ moonshotai/Kimi-Audio-7B-Instruct · Kimi-Audio is an open-source audio foundation model built for audio understanding, generation, and conversation. It supports tasks like ASR, audio Q&A, captioning, and emotion recognition, trained on 13M+ hours of audio with a novel architecture and low-latency inference.

⭕ Anthropic/values-in-the-wild · Anthropic shares a dataset of 3,307 values expressed by Claude across real-world conversations, using a privacy-preserving method with no human content access. It supports research into how AI systems demonstrate values in practice and offers a structured taxonomy for interdisciplinary study.

⭕ OpenGVLab/InternVL-Data · The InternVL3 Open Dataset supports multimodal research across image, text, and video understanding. It includes data from open sources, synthesized content, and the web. Initial releases cover InternVL2.5 and InternVL3 SFT data, with full uploads and distribution details coming over the next few weeks.

⭕ How to Create a Custom Model Context Protocol (MCP) Client Using Gemini? This guide walks you through creating a custom Model Context Protocol (MCP) client using Gemini 2.0 Flash. It shows how to connect Gemini with MCP servers, configure tools, handle queries, and interact via command line, enabling real-time AI responses with tool execution.

⭕ Devin AI Introduces DeepWiki: A New AI-Powered Interface to Understand GitHub Repositories. Devin AI has introduced DeepWiki, a free tool that auto-generates interactive, wiki-style documentation for any GitHub repository. Using their DeepResearch agent, it offers project summaries, architecture diagrams, and module insights, streamlining code understanding without installation or setup. Just swap github.com with deepwiki.com to start.

⭕ Adding Training Noise To Improve Detections In Transformers: New techniques like DN-DETR and DINO improve object detection in vision transformers by adding noise to ground truth boxes during training, which stabilizes learning, reduces reliance on complex matching algorithms, and accelerates convergence, yielding stronger, faster models for 2D, 3D, and temporal detection tasks.

⭕ A closer look at BigQuery DataFrames 2.0: Google has introduced BigQuery DataFrames 2.0, bringing multimodal data processing, across text, images, and audio, into scalable Python workflows, allowing data scientists to use familiar Pandas-like syntax while handling massive datasets with generative AI, vector search, and natural language-powered SQL built directly into BigQuery.

⭕ Graph Neural Networks Part 4: Teaching Models to Connect the Dots. Graph neural networks are now powering smarter link prediction by combining simple heuristics like Jaccard and Adamic-Adar with deep models like VGAE, which learn node relationships directly from graph structures, allowing systems to predict connections in social networks, recommend items, and map interactions more accurately than traditional rule-based approaches.

⭕ Sharing new DORA research for gen AI in software development: DORA's latest report reveals how generative AI is already reshaping software development, with 76% of technologists using it in daily work and clear links to productivity, code quality, and review speed, backed by five actionable strategies to drive adoption, empower teams, and responsibly integrate AI across organizations.

⭕ Customize Amazon Nova models to improve tool usage: Amazon Nova models can now be customized for precise tool usage through supervised fine-tuning in Amazon Bedrock, enabling more accurate argument extraction, better integration with APIs, and improved performance in agentic workflows, allowing developers to build smarter, more responsive AI systems tailored to real-world decision-making needs.

⭕ Building Fully Autonomous Data Analysis Pipelines with the PraisonAI Agent Framework: A Coding Implementation. PraisonAI Agents, powered by Google Gemini, enable fully autonomous data analysis pipelines through natural-language prompts, letting users load, filter, summarize, group, pivot, and export datasets without writing Pandas code, while self-reflection and verbose logging ensure transparency, traceability, and human-readable reasoning at each step.

⭕ Implementing Persistent Memory Using a Local Knowledge Graph in Claude Desktop: Implementing a local knowledge graph in Claude Desktop enables persistent memory across chats, allowing it to recall user identity, preferences, and past context using MCP tools, resulting in more personalized, consistent interactions without repeating details, all through a simple setup using Node.js and configurable settings.

⭕ Choose the Right One: Evaluating Topic Models for Business Intelligence: In this tutorial, bigram topic models like BERTopic and FASTopic are evaluated for classifying customer emails, with metrics such as semantic coherence, normalized PMI, and uniqueness guiding model selection, helping businesses prioritize responses, reduce support time, and make more effective decisions based on nuanced topic insights.

⭕ Evaluate Amazon Bedrock Agents with Ragas and LLM-as-a-judge: Amazon Bedrock Agents can now be systematically evaluated using Ragas and LLM-as-a-judge techniques, allowing developers to assess RAG, text-to-SQL, and chain-of-thought performance through structured metrics, improving transparency, safety, and optimization in single or multi-agent workflows, with results visualized in Langfuse for deeper insights.