hessian.AI’s Post

Exciting to see the impact of products created by QuantumBlack, AI by McKinsey, and Kedro now has almost ~17M downloads and 10K stars on GitHub. Kedro is an open-source Python framework that uses software engineering best practices to help you build production-ready data science code that is reproducible, maintainable, and modular, Market studies indicate that ~90% of data science projects do not make it into production and usage in the field, suggesting the last 5 years have been defined more by proof-of-concept AI than operationalized value. Kedro was created by Yetunde Dada and the team within QuantumBlack Labs, our AI/ML innovation and software development hub with 250 technologists building tools to support the work of QB's more than 1300 data scientists across 50 locations. In 2022, McKinsey donated Kedro to the Linux Foundation, AI & Data. #QuantumBlack #AIbyMcKinsey #MachineLearning #Kedro

🚀 Unleashing the Power of Structured Outputs with LLMs! 🧠 Ever tried integrating Large Language Models (LLMs) with your software systems and hit a wall due to inconsistent responses? 🤔 The brilliance of LLMs lies in their contextually rich outputs, but their variability can be a major stumbling block when structured data is crucial. 🔍 Curious about how to tackle this challenge? Dive into how Pydantic and LangChain can transform your LLM outputs into reliable, structured data! 🔎 The Story Behind the Solution: In the world of AI, where LLMs excel at generating human-like responses, a key challenge emerges—consistency. Imagine you need structured outputs like JSON for seamless software integration, but LLM responses are unpredictable. 🤯 Enter Pydantic: The game-changer in data validation! By defining custom object types with Pydantic, you can enforce a consistent structure in LLM responses. Think of it as crafting a blueprint for your data, ensuring that every piece of information fits perfectly into your system's needs. 🏗️ Introducing PydanticOutputParser: Within LangChain, this powerful tool leverages Pydantic to parse JSON responses effectively. 📊 For instance, generating a story with character names and places can be simplified using Pydantic to create structured data models. 📝 🔄 Here’s how it works: - Define your output schema using Pydantic’s BaseModel. - Set up a parser and prompt template with LangChain. - Get structured, reliable outputs from LLMs every time! Example Use Case: Creating a story with character names and places? Pydantic ensures that the output is always in the format you need, overcoming the challenge of variable response formats. 🌟 🔗 Follow me for more insights into leveraging AI for seamless integration and share this post with your network to help others solve similar challenges! #AI #MachineLearning #Pydantic #LangChain #DataScience #LLM #TechInnovation #StructuredData #Python #SoftwareDevelopment

PandasAI (YC W24) just secured $1.1M in Pre-Seed funding to revolutionize data analysis with AI. This German startup is blending the power of generative AI with the popular Python library, pandas, making data querying easier and more effective. Here’s why this is game-changing: 1. AI-Enhanced Analysis: PandasAI turns complex data questions into simple, conversational queries. 2. Open-Source Momentum: With 8,600+ stars on GitHub, it's already gaining traction among Fortune 500 companies. 3. Scalable Vision: The funding will help founder Gabriele Venturi scale up, bringing advanced data insights to decision-makers faster. Gabriele Venturi, Adam Shuaib, PhD What’s your biggest data analysis challenge right now? https://lnkd.in/dRvTe47y #AI #DataAnalysis #StartupInnovation #mondayfunding

🔬 𝐆𝐞𝐦𝐦𝐚-𝟐 𝐯𝐬 𝐋𝐋𝐚𝐌𝐀-𝟑 Building on my recent Gemma-2 post, I did a comparison of Gemma-2 and LLaMA-3 using Tune Studio's playground. Here are some observations. 𝐓𝐞𝐬𝐭 𝐐𝐮𝐞𝐬𝐭𝐢𝐨𝐧𝐬: 1. Explain quantum entanglement to a high school student 2. Implement the bubble sort algorithm in Python 3. Logical reasoning: "If all A are B, and some B are C, what can we conclude about A and C?" 4. Compose a haiku about artificial intelligence 5. Discuss ethical implications of AI in healthcare decision-making 𝐊𝐞𝐲 𝐎𝐛𝐬𝐞𝐫𝐯𝐚𝐭𝐢𝐨𝐧𝐬: 1. Conceptual Explanations:  - Gemma-2: Offers detailed, step-by-step explanations with multiple analogies  - LLaMA-3: Provides concise overviews with focus on key principles 2. Code Implementation:  - Gemma-2: Produces optimized, well-commented code with additional explanations  - LLaMA-3: Delivers concise, functional code with basic usage examples 3. Logical Reasoning:  - Both models demonstrate strong capabilities  - Gemma-2: More concise with immediate practical examples  - LLaMA-3: Provides step-by-step logical breakdown 4. Creative Tasks:  - Both models generate thematically appropriate haikus  - Gemma-2 focuses on learning and human aspirations  - LLaMA-3 emphasizes technological progress 5. Ethical Analysis:  - Gemma-2: In-depth discussion of specific ethical issues  - LLaMA-3: Structured overview with clear categorization of benefits, implications, and mitigations 𝐀𝐝𝐝𝐢𝐭𝐢𝐨𝐧𝐚𝐥 𝐈𝐧𝐬𝐢𝐠𝐡𝐭𝐬: - Response Structure: LLaMA-3 consistently uses more structured formats (bullet points, sections) - Depth vs. Conciseness: Gemma-2 tends towards detailed explanations, while LLaMA-3 offers more succinct responses - Practical Application: Gemma-2 often includes more real-world examples and analogies 🎥 𝐅𝐮𝐥𝐥 𝐜𝐨𝐦𝐩𝐚𝐫𝐢𝐬𝐨𝐧 𝐯𝐢𝐝𝐞𝐨: https://lnkd.in/dE5f9VGC 🚀 Do you wanna try this experiment? You can test these models yourself on Tune Studio Playground. We support all the open source models! 👨💻 Try live here: https://studio.tune.app PS: Models used: gemma-2-27b-it and Meta-Llama-3-70B-Instruct #AIModelComparison #MachineLearning #Gemma2 #LLaMA3 #TuneStudio

🚀 **Exciting Milestone Achieved!** 🚀 I’m thrilled to announce that I’ve successfully migrated my trading strategy from MetaTrader to **TradeLocker Studio**, a next-gen beta feature that leverages the **Backtrader Python framework** for algorithmic trading! 🎯 This transition has unlocked a whole new world of flexibility and customization for my strategies, and the results speak for themselves! 💡 With **3.36% ROI** and an **annualized ROI of 41.18%**, this migration has been nothing short of a game-changer. While there are still a few limitations—particularly with resolution and backtest time frames—I'm confident these will improve as I continue to optimize. 💻🔧 But there’s more to be excited about in the realm of AI and trading: **OpenAI O1 Preview** is on the horizon, offering even more possibilities for integrating **Agentic AI** into our workflows. Agentic AI—autonomous systems that can make decisions—has the potential to radically enhance how we work by handling complex tasks and enabling better decision-making, faster. I truly believe the future of work will be shaped by **Human-AI collaboration** 🤖🤝. Imagine AI automating routine processes while humans focus on the creative, strategic side—pushing boundaries and unlocking new potentials. In fields like **algorithmic trading**, this synergy can lead to faster, more informed decisions with powerful, data-driven insights. The future of work isn't about replacing human intelligence but **augmenting it**. The collaboration between Agentic AI and humans can reshape industries by combining the best of both worlds: AI’s speed and precision with human intuition and creativity. It’s been an incredible journey so far, and I can't wait to explore more with **TradeLocker Studio** and beyond! #AlgorithmicTrading #Python #TradeLockerStudio #Backtrader #OpenAI #AgenticAI #Fintech #TradingStrategy #FutureOfWork

𝐄𝐱𝐩𝐥𝐨𝐫𝐢𝐧𝐠 𝐋𝐚𝐧𝐠𝐂𝐡𝐚𝐢𝐧 𝐀𝐠𝐞𝐧𝐭𝐬 𝐟𝐨𝐫 𝐀𝐝𝐯𝐚𝐧𝐜𝐞𝐝 𝐀𝐈 𝐀𝐩𝐩𝐥𝐢𝐜𝐚𝐭𝐢𝐨𝐧𝐬! I've been diving into LangChain's 𝗔𝗴𝗲𝗻𝘁𝘀 and their powerful built-in tools like 𝗗𝘂𝗰𝗸𝗗𝘂𝗰𝗸𝗚𝗼 search and 𝗪𝗶𝗸𝗶𝗽𝗲𝗱𝗶𝗮 for dynamic and enriched responses! LangChain offers a flexible, open way to combine multiple tools, making it a go-to for creating highly responsive AI solutions. 𝐎𝐯𝐞𝐫𝐯𝐢𝐞𝐰 𝐨𝐟 𝐭𝐡𝐞 𝐏𝐫𝐨𝐜𝐞𝐬𝐬: 1. 𝗦𝗲𝘁𝘁𝗶𝗻𝗴 𝘂𝗽 𝗟𝗮𝗻𝗴𝗖𝗵𝗮𝗶𝗻 𝘄𝗶𝘁𝗵 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝘃𝗲 𝗔𝗜: After setting up with langchain-google-genai and wikipedia packages, I initialized 𝗚𝗼𝗼𝗴𝗹𝗲’𝘀 𝗚𝗲𝗺𝗶𝗻𝗶-𝟭.𝟱 model for natural responses with a balanced temperature for reliable outputs. 2. 𝐔𝐬𝐢𝐧𝐠 𝐁𝐮𝐢𝐥𝐭-𝐈𝐧 𝐓𝐨𝐨𝐥𝐬: With the ChatGoogleGenerativeAI model, I explored 𝗪𝗶𝗸𝗶𝗽𝗲𝗱𝗶𝗮 𝗮𝗻𝗱 𝗟𝗟𝗠 𝗠𝗮𝘁𝗵 𝘁𝗼𝗼𝗹𝘀 to fetch information and solve mathematical queries seamlessly, like calculating "What is 25% of 300?"—showcasing how adaptable LangChain can be for various informational needs. 3. 𝐂𝐮𝐬𝐭𝐨𝐦 𝐏𝐲𝐭𝐡𝐨𝐧 𝐀𝐠𝐞𝐧𝐭𝐬: For specific tasks, I created a 𝗣𝘆𝘁𝗵𝗼𝗻 𝗮𝗴𝗲𝗻𝘁 to sort a list of customer names by last and first names, leveraging Python’s REPL environment. The agent’s toolchain adapts to sorting requests, highlighting LangChain's flexibility in handling customized processing requirements. 4. 𝗗𝗲𝗳𝗶𝗻𝗶𝗻𝗴 𝗖𝘂𝘀𝘁𝗼𝗺 𝗧𝗼𝗼𝗹𝘀: I created a custom tool to provide the current date, offering a quick answer to "𝘞𝘩𝘢𝘵'𝘴 𝘵𝘩𝘦 𝘥𝘢𝘵𝘦 𝘵𝘰𝘥𝘢𝘺?"—a great demonstration of how easy it is to expand LangChain with unique tools for enhanced utility. 𝗧𝗮𝗸𝗲𝗮𝘄𝗮𝘆: LangChain’s agent tools provide a dynamic environment to blend LLM capabilities with custom logic. While agents may occasionally need reruns for accuracy (agents are still evolving!), the versatility and scope of LangChain make it perfect for advanced AI applications. #AI #LangChain #MachineLearning #ArtificialIntelligence #Python #DataScience #GoogleAPI Sir Ameen Alam Sir Muhammad Qasim

Explore the cutting-edge capabilities of AI pattern creation with Daniel Miessler's Fabric framework. Check out the GitHub repository Fabric @https://lnkd.in/gcVbBhwD to see how you can start building custom AI patterns today. Perfect for developers interested in advancing their skills in AI-driven technology!' Follow Daniel Miessler #FabricAI #TechCommunity #SoftwareDevelopment #MachineLearning https://lnkd.in/gcVbBhwD

Har Har Mahadev! 🙏🙏🙏 thanks to Krish Naik iNeuron.ai Overcoming Categorical Data Issue in Data Science: A Case for SMOTE Real-life data is often plagued with a class imbalance problem, a common problem that is predominant in fraud detection, medical diagnosis, churn prediction, among other domains. Model performance suffers considerably due to the fact that machine learning algorithms tend to be biased toward the majority class. What are the Ways to Solve this? Fortunately, there is a method called SMOTE (Synthetic Minority Over-sampling Technique). SMOTE creates new minority class observations that allow the distribution of classes within the data set to be more even, resulting in a better predictive model. The way SMOTE works is as follows: Creation of Synthetic Data: Despite being a very concentrated approach, the SMOTE technique eschews blind duplication of existing data where it interpolates the space between two (or more) samples from the minority class to generate new observations. Enhanced Model Effectiveness: After the distribution of classes has been addressed, classifiers such as Logistic Regression and decision trees will be able to learn more effectively how to separate the classes. Implementation of SMOTE: It is straightforward to use with packages such as imbalanced-learn for Python: How to Apply SMOTE: It’s easy to implement using libraries like imbalanced-learn in Python: from imblearn.over_sampling import SMOTE smote = SMOTE() X_resampled, y_resampled = smote.fit_resample(X, y) Key Points: SMOTE works best when applied before training. Combine it with techniques like Random Undersampling for even better results (SMOTEENN). Use performance metrics like precision, recall, F1-score, and AUC-ROC to evaluate model performance on imbalanced data. Balancing your dataset with SMOTE can make a significant difference in the accuracy and fairness of your predictions! #DataScience #MachineLearning #ImbalancedData #SMOTE #AI #DataScienceTips

🎬 𝐄𝐱𝐜𝐢𝐭𝐞𝐝 𝐭𝐨 𝐒𝐡𝐚𝐫𝐞 𝐌𝐲 𝐋𝐚𝐭𝐞𝐬𝐭 𝐏𝐫𝐨𝐣𝐞𝐜𝐭: 𝐌𝐨𝐯𝐢𝐞 𝐑𝐞𝐜𝐨𝐦𝐦𝐞𝐧𝐝𝐚𝐭𝐢𝐨𝐧 𝐒𝐲𝐬𝐭𝐞𝐦! 🎬 I’m excited to announce the completion of my latest project—a Movie Recommendation System based on content-based prediction. This project was an interesting journey into the world of data science and machine learning, where I delved deep into the intricacies of recommendation algorithms. 🔍 𝐊𝐞𝐲 𝐇𝐢𝐠𝐡𝐥𝐢𝐠𝐡𝐭𝐬: 𝐂𝐨𝐧𝐭𝐞𝐧𝐭-𝐁𝐚𝐬𝐞𝐝 𝐅𝐢𝐥𝐭𝐞𝐫𝐢𝐧𝐠: Utilized various features such as genre, director, cast, and plot summaries to recommend movies similar to ones users have enjoyed in the past. 𝐍𝐚𝐭𝐮𝐫𝐚𝐥 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐏𝐫𝐨𝐜𝐞𝐬𝐬𝐢𝐧𝐠: Employed NLP techniques to process and analyze movie descriptions for better feature extraction. 𝐔𝐬𝐞𝐫 𝐄𝐱𝐩𝐞𝐫𝐢𝐞𝐧𝐜𝐞: Designed with a user-friendly interface to provide seamless movie recommendations tailored to individual preferences. This project not only enhanced my technical skills in 𝑃𝑦𝑡ℎ𝑜𝑛, 𝑃𝑎𝑛𝑑𝑎𝑠, 𝑎𝑛𝑑 𝑆𝑐𝑖𝑘𝑖𝑡-𝐿𝑒𝑎𝑟𝑛 but also honed my understanding of real-world applications of machine learning. Feel free to check out my project and share your thoughts! 🚀 𝐀𝐩𝐩 𝐥𝐢𝐧𝐤 : https://lnkd.in/grYzqHu7 𝐆𝐢𝐭𝐡𝐮𝐛 : https://lnkd.in/gc9NS6qA

🌟 𝐑𝐞𝐜𝐞𝐧𝐭𝐥𝐲 𝐰𝐫𝐚𝐩𝐩𝐞𝐝 𝐮𝐩 𝐚𝐧 𝐢𝐧𝐜𝐫𝐞𝐝𝐢𝐛𝐥𝐞 𝐡𝐚𝐜𝐤𝐚𝐭𝐡𝐨𝐧 𝐞𝐱𝐩𝐞𝐫𝐢𝐞𝐧𝐜𝐞! 🌟 I'm excited to share my journey at 𝐋𝐥𝐚𝐦𝐚 𝐈𝐦𝐩𝐚𝐜𝐭 𝐇𝐚𝐜𝐤𝐚𝐭𝐡𝐨𝐧, organized by lablab.ai Meta MindsDB LlamaIndex Together AI AutogenAI Snowflake Restack Featherless AI AI Alliance.🏆 Our team worked on an exciting project called 𝐋𝐥𝐚𝐦𝐚𝐄𝐯𝐚𝐥, 𝐚 𝐐𝐮𝐢𝐜𝐤 𝐄𝐯𝐚𝐥𝐮𝐚𝐭𝐢𝐨𝐧 𝐃𝐚𝐬𝐡𝐛𝐨𝐚𝐫𝐝 which is a streamlined tool designed to simplify the evaluation of Llama model outputs. We aimed to make model evaluation faster and more intuitive for developers, with real-time performance comparisons across various tasks. 🤖 💻 My role was on the front-end development, where I used 𝐒𝐭𝐫𝐞𝐚𝐦𝐥𝐢𝐭 with Python to design and build the 𝐖𝐞𝐛 𝐔𝐈, making the tool user-friendly and accessible (without knowing anything about Streamlit and Llama). Additionally, I prepared the presentation slides and project description to effectively showcase our work. 🙌 Special thanks to my amazing teammates: Paraskevi Kivroglou (Team Leader) and Mayank Varshney. Working together has been an incredible learning experience, and I'm grateful for the opportunity to combine our skills and knowledge to create something impactful. 🔗 Project: https://lnkd.in/dDDK_Wre Thanks to this experience, I feel even more motivated to dive into AI and look forward to building impactful solutions. Here's to more projects, more learning, and more growth! 💡 #Hackathon #AI #Teamwork #LlamaEval #TogetherWeBuild #Learning #llamahackathon #lablabai #LlamaImpact #Meta #TogetherAI #MachineLearning #Growth