Oana O.’s Post

Uber’s Michelangelo was an end-to-end ML system. It was a perfect fit for the early users. But… As ML teams scaled, more and more found gaps. In the 8-stage ML pipeline (even with a 90% chance of success at each stage), the all-in-one solution may not meet the expectations. → Prioritize adaptability. While end-to-end systems offer a good range of features, point solutions might be better equipped for more specific ML challenges or requirements. h/t to Michael Del Balso from Tecton for the insight. — (link to the full episode in the comments) #ML #MLOps #MLSystem

Excited to share what we have been working on for the last few months 🚀 We are building the AI that helps build software for everyone! Lutra writes the code, interconnects with your apps, and produces AI assistants for you -- all just from English. https://lnkd.in/gdW-bviu Earlier this year, I was surprised at the ability of models like #GPT4 to reason and code -- coding unlocks a world of possibilities for AI models to interact with the world. But, the sandboxes (like #codeinterpreter) were severely limited as they are unable to connect externally. Interacting with our own data and apps was even more challenging as we still have to figure out all the API integrations - and where to run this securely. We set out to make all of this easy with https://lutra.ai Lutra comes integrated in an AI-native way, optimized for AI models. For example, if you wanted all your emails classified by importance, you simply tell Lutra -- "Go through all my inbox emails, and classify them by importance" -- and it works! Essentially, this is AI generated software. We are on the precipice of a new era of computing where software is no longer one-size-fits-all, and instead, software is made on the fly for our needs. I am thrilled and grateful to be able to work again with some of the best people I know including Vijay Vasudevan, Bo Zhi See, Joshua Newman, Jaran (Jerry) Charumilind, and Marc Rasi. I am also thankful for the support from all our investors and angels. More at https://lnkd.in/gdW-bviu Follow me on https://lnkd.in/gHjbxgHX for updates!

At Lutra AI we're building AI to help anyone build software. Describe what you want in English and Lutra builds a code-first AI workflow that you can immediately run, iterate on (also in English!), share, and run on a schedule. We've designed it to be secure, reliable, transparent, and maintainable, even for those who aren't experts in software engineering. I already use Lutra daily to keep up with important news in the explosive field of Generative AI, automatically submit reimbursements for receipts sent to my email, and a host of other time-consuming flows. We're busy building many more integrations with tools we all use in our personal and work lives. Check out more at lutra.ai and sign up for our waitlist! 🚀

The era of LLMs is a chance to start over & avoid the tool sprawl that plagued ML ops. Freeplay's end-to-end platform empowers dev & product teams to collaborate on optimizing LLM-powered products that drive real outcomes for customers - without the headache of disjointed tools. Learn more & get started at https://freeplay.ai

Have been using #LangSmith, a groundbreaking beta project from LangChain and recommend you signing up for the beta to avail of its vast array of current (and growing !!) features. Not only does it allow you to debug, test, evaluate, and monitor chains and intelligent agents built on any #llm framework and seamlessly integrate with #langchain , but it does so in an intuitive, visually appealing interface. Features such as "Open in Playground" allows you to modify the #prompt , re-run it and visualize the changes as many times as needed, all from within the same interface; depicting the exact sequence of events as they occur; exact number of #tokens used, time to execute each step etc. are an immense help in the testing, evaluation, #observability of enterprise grade #generativeai applications... Snippet below from one of my runs in the platform... #promptengineering #genai #llmops

> Standardization. Stability. Reusability. As I reflect on my chat with Piotr Niedzwiedz & Aurimas Griciūnas, these three words are a pretty reasonable way to characterize the goals of any "centralization" or "platform" effort, or so I think! So I have an exercise, for anyone building/wanting to build a platform, can what you're doing be described with these three words? or is something missing? or is there a gap in your efforts? For example, let's see how Hamilton fits with these three words: Standardization: Hamilton helps standardize python code written by way of its opinionation. That way anyone can pick it up, understand what's going on. Anyone on the team has to follow this convention. Stability: From a testing perspective, Hamilton ensure code can be well tested and documented, leading to stable changes & updates. In addition, there is standardization, so changes made to code can be done in a very prescribed manner. This means that as people come and go to projects, there is stability within the project. Reusability: Hamilton forces curation of code into modules. So by definition you get reusability quite easily. Teams can centralize code and therefore share it and make use of it more easily than before, as there's a standard way to go about consuming it. Thoughts? How did you get on? Do you agree? Disagree?

MLOps 101: Experiment Tracking MLflow is a popular tool for experiment tracking and model management, but it's not the only option available. Although the specific tools may differ, the fundamental concepts of experiment tracking remain largely the same across different solutions. #MLOps #MLflow #MachineLearning #MLEngineering

Amazing work from Jiquan Ngiam, Bo Zhi See, Vijay Vasudevan and the whole Lutra AI team on launching today! I was lucky enough to get to test this out pre-launch, and it's absolutely incredible to see what the team has been able to put together in such a short window of time. They are helping make my job a lot easier through automating several pieces of what I do on a regular basis, and there's several different applications to what they can do - I highly recommend you check them out. #ai #googlecloud

Ever wondered how to make your ML projects smoother and more organised? For me MLOps is the answer. It's like having a magical helper that makes deploying your awesome models way easier. Instead of feeling lost in a maze when you try to put your model into action, MLOps steps in to make everything super streamlined. It handles stuff like version controlling, productionalisation, continuous integration, and making sure everything works together seamlessly. Bottom line? MLOps saves you time and headaches, making your model's journey from idea to reality a breeze! ✨ #MLOps #MachineLearning #SmoothSailing

David DeSanto, Chief Product Officer at GitLab, predicts a paradigm shift in 2024, with AI at the forefront. Explore GitLab's 2023 Global DevSecOps Report for a sneak peek into the evolving landscape of software development toolchains. While the accelerated integration of AI tools may bring challenges, such as biased outputs, DeSanto envisions this phase driving the establishment of robust ethical guidelines. Read more about the transformative journey ahead! 📰 https://brnw.ch/21wGhVR #AIInnovation #SoftwareDevelopment #TechFuture