Mariia Sandrikova

I'm excited to announce that our talk on the Agent-based AI Framework for ROS 2 has been accepted for ROSCon 2024 🐢! In recent years, there have been significant advancements in Visual-Language-Action (VLA) models. However, these advancements have not yet enabled these models to independently perform real-life robotic tasks. Rather than solely pursuing cutting-edge AI models aimed at complete transformations, we prioritize practical applications that deliver clear benefits in real-world scenarios. Our approach emphasizes seamless integrations and the creation of robust tools to effectively support everyday robotic tasks. If you're interested or have any questions, please feel free to reach out! #ros #ros2 #roscon #robotics #vla

8

🤟 Exciting news! My talk "Representation is King 👑: The Journey to Quality Dialog Embeddings" that I presented at EuroPython is now available for viewing! I dived into how sentence embeddings can enhance the classification of dialogs, showcasing the use of top-performing small-sized sentence transformers. The talk is packed with intuitive, hands-on examples using the SentenceTransformers framework to fine-tune models with labeled and unlabeled data 👐. We leverage similar techniques at Salted CX to create innovative analytical contact center solutions 😎

39

So I became fascinated by C++ coroutines late last week, and have spent a good portion of the weekend figuring out how they work. The TL;DR of my discoveries so far is that C++20 coroutines: 1) follow the best and the worst of C++ practices, and 2) are a really clever way to decouple stackless functions from their execution model. For (1), well, it's C++. No crisp & clean syntax, lots of weird compilation errors, and even weirder runtime errors. I hated them with passion for the first several hours. But then they grew on me, because the flip side is also true. The flip side is backwards compatibility and "only pay for what you use". And introducing a totally new concept — the stackless execution model, effectively, [proto-] green threads — into a mature and complex language is one hell of a pain! For (2), I do believe it's a genius solution for the language to separate stackless functions from their execution model. And C++20 did exactly that. There is literally no event loop in C++20! The claim (as I understand it) is that a tiny microcontroller may want to use C++20 coroutines, and the authors of the Standard are very explicit about not constraining possible runtimes. All that C++20 coroutines do is introduce stackless functions into the language as first-class citizens. And yes, stackless immediately implies resumable. So it is one step about the most-clever-possible syntactic sugar over `setjmp()`. Full one step. But not a millimeter more than this one step. Whether you "execute" your coroutines in a dedicated single-threaded event loop, or a thread-pool-based event loop, or by individual steps via some timer event or even a network packet — that is entirely your call. The stackless functions are there, `co_await` (as well as `co_yield` and `co_return`) is there, no callback hell is needed, and all the RAII and scoped initialization/ownership model is fully intact. Now, I'm still elbows deep in my experiments, but here's the first big LinkedIn-worthy question. Why don't we write our C++20 [unit] tests as corourines? Say, you want to limit some test to five seconds. And fail it as "timed out" within these five seconds. Before coroutines, doing this without instrumenting the very code of the test involved `fork()`-ing the process, effectively relying on the operating system's primitives to ensure proper resource deallocation. With C++20 coroutines, as long as the slow test is slow on non-blocking I/O (i.e. not in an infinite loop, but on something that `co_await`-s stuff), it is possible to time-box each individual test to some five seconds very much organically. Because a coroutine frame, unlike a stack frame, can be `.destroy()`-ed safely and cleanly. So, are we doing it yet? If no, why not? If yes, what would be some next-gen `gtest` to look into? Also, I'll be writing some code and blog posts on C++20 coroutines soon. If you are reading this text and are an expert, DM me pls — I'd love to pick your brain!

19

4 Comments

Automation of recruitment processes is a high value prize for many companies. One of the biggest roadblocks is a higly irregular structure of CVs combined with massive information density. Person entire career is squished to 1 pdf page. Under these conditions requirements to ML models on quality of data extraction are extremely high. Having open source dataset of CVs can give a massive boost to research effort and help bring to production quality existing models Great job!

9

FlexAttention generated a lot of buzz in the PyTorch Community this week. I want to explain why this API is an important milestone for the compiler team. FlexAttention is the first major Eager-like API powered by torch.compile() under the hood.  It opens up a new venue for deploying torch.compile() with few hassles to end-users. Since we compile a library code that is co-designed with the compiler, compile-backed Eager APIs avoid most scenarios of compiling user codes that may require expert attention (e.g., fixing graph breaks, recompilation, composability, or performance problems). The compiler team is exploiting more cases of "small-scope" compilation (also known as Higher-Order Kernels, HOK) to increase the adoption of torch.compile(), especially among power ML users who are also performance experts, enjoy the hackability of PyTorch Eager, and do not want the compiler to take control of the whole world. Stay tuned, and hope to meet you in PTC'24, 09/18-09/19, in SF (registration open, https://lnkd.in/eH8a-3J2). https://lnkd.in/eJqUda4N

334

2 Comments

Today I have applied for the participation in Junction 2024 - back offline hackathon. It’s me, Igor Vernichenko and Aarni Pavlidi who will try to come up with something super exciting for the second time. Or have just fun as we did last time. #junction2024 #helsinkievents #backoffline #hackathon2024 #junction #hackathon

21

3 Comments

🚀 New Project: Spotify Track Downloader with Enhanced Metadata! 🎧 I'm excited to share a project I've been working on—my very own Spotify Track Downloader web app! This was a challenging and rewarding experience, where I got to dive deep into full-stack development, API integration, and audio processing. ✨ What’s it all about? This web app allows users to download Spotify tracks as MP3 files, complete with enhanced metadata such as song name, artist name, and album art. From OAuth 2.0 authentication for secure API access to embedding album art and track info into the MP3 files, this project really put my skills to the test and was a great way to learn about real-world applications. 🎥 Check out the demo: I've put together a quick demo video to showcase how it works. You can watch it below or at this link! https://lnkd.in/eWCc5Dqx 🔧 Tech Stack: • 🐍 Python & Flask for the backend. • 🎨 HTML, CSS, JavaScript, Bootstrap, and jQuery for the frontend. • 🎵 Spotify Web API (for metadata and 30-second track preview) & YouTube Data API (for full tracks). • 🔐 OAuth 2.0 for secure authentication. • 📚 Libraries like Spotipy, yt-dlp, and Mutagen for API interactions and metadata management. 💻 Check out my code here: https://lnkd.in/e-_NZMcz This project was a great way to apply my coding skills to a real-world application, and I’m proud of the results. If you’re interested in learning more, feel free to check out the full project in my LinkedIn profile! Disclaimer: "Although I have the capabilities to host this application live, I have chosen not to do so due to legal considerations. This project and demo are for educational purposes only, showcasing my skills in coding and API integration. The Spotify Track Downloader is not intended for public use, and I do not intend to profit from it. Please note that downloading copyrighted content without permission from the rights holder is illegal and violates Spotify's terms of service. Always respect copyright laws and the rights of content creators."

19

3 Comments

🎓 I'm proud to announce my graduation with distinction from the Elephant AI PRO™ course, deepening my expertise in AI for management. Special thanks to Maria Parysz for the honor of receiving "The Most Impactful Initiative" badge – a recognition that truly motivates me. Big thanks also to the wonderful team at Elephant AI! Paulina Ołtusek, Dominika oraz Maja Kostyrka. With substantial experience in Mendix, this course has equipped me to bridge the gap between low-code development and AI, opening a world of possibilities. I'm excited to blend these domains to innovate and streamline business processes. Watch this space as I merge the agility of low-code with the intelligence of AI to create smarter solutions for the future. #ArtificialIntelligence #LowCodeDevelopment #Mendix #AIforManagers #ElephantAI #Innovation #FutureOfTech

51

6 Comments

We just released Bring Your Own Model in Nightly! This feature lets you run your AI browser assistant with on-device open models (e.g. Llama 3 8B, Mistral 7B, Phi 3). Chat, query webpages, use smart tools and more - without your data leaving your control. Easy setup with tools like Ollama. More info in the blogpost linked below.

15

Meet Jordi A., a 4th-year PhD student in Computer Science at the University of Edinburgh. This summer, he joined Meta's Fundamental AI Research (FAIR) team as a Research Scientist Intern. "I'm working on large language models and code generation. I'm still astounded by the massive number of GPUs and the level of the research collaborators I have access to at Meta," Jordi explained. "I'm investigating how to get AI models to have a better understanding of code by using information obtained when executing programs." Jordi was drawn to Meta because of FAIR's groundbreaking open science and ambitious plans for Artificial General Intelligence (AGI). "I'm so excited to be part of a lab working on the frontiers of AI research," he said. The best advice he received this summer? "Make sure you meet all the relevant people — there's always something to learn from other's work. It's easier to limit yourself to your immediate peers but going the extra mile to meet researchers and interns from other teams or locations has been totally worth it." After completing his internship, Jordi plans to finish his PhD and continue researching AI for code generation. "There is so much more that AI can do for us — this is just the beginning.” #LifeAtMeta #AI #AIResearch #MetaIntern

14

1 Comment

🔥 Excited to share that I was a guest on the latest "AWS in Russian" Podcast episode about hosting LLM in the cloud. Big thanks to Viktor Vedmich for inviting me. We discussed deploying large language models (LLMs) on AWS: • Amazon Bedrock • SageMaker • ECS and EC2 • AWS Lambda Learn where to start, how to choose the right solution, and optimize costs. Listen here: https://lnkd.in/gv_wZF4G #AWS #MachineLearning #AI #Podcast #Cloud

51

6 Comments

🌐 Announcing the Blockchain Compression Project 🚀 I am pleased to present my latest research-driven project, Blockchain Compression Project, which aims to address the growing challenges of blockchain data storage and scalability. This initiative incorporates advanced techniques such as: - Data compression to minimize the storage footprint, - Merkle Trees for efficient transaction verification and data integrity, - State Delta Management to track and store only changes, reducing unnecessary duplication, - ZK-SNARKs to ensure cryptographic privacy and verification without compromising security. The project is designed to optimize the storage of blockchain data while maintaining robustness, security, and scalability, making it a valuable contribution to the ongoing development of efficient blockchain systems. I invite researchers, developers, and blockchain professionals to explore the project and contribute to its evolution: https://lnkd.in/gW2qzTU6 #BlockchainResearch #DataCompression #Cryptography #ZKSNARK #MerkleTree #Scalability #OpenSource

3

I have a lot of knowledge and interests, but I really want to focus on making LLMs smaller for specific tasks. The models would still do all the same things as the original LLM, but they would be better at certain things. They would also be cheaper and take up less space to run. These models would be able to generalize better than older models, meaning they could handle more situations. But they would still be more affordable compared to the original LLM. I am pleased to note that the 14 million parameter model ( https://lnkd.in/ef_9rxc7 ) that I published on Hugging Face has attracted over 170 downloads within a week. Machine learning specialists in the field of e-commerce require solutions that can be immediately applied to everyday tasks.

13

Roofline Analysis of Apple Silicon GPUs using MLX I wrote a tutorial (https://lnkd.in/d8N6_8aU) on Roofline analysis and its application to Apple silicon GPU kernel development, a method traditionally used by CUDA kernel developers. A big thank you to Angelos Katharopoulos and Awni Hannun, the creators of MLX, for their invaluable help in making this tutorial possible. The inspiration came from friends who asked why Recurrent Drafting achieves higher speedups on the M2 Ultra compared to the M1 Max after reading my previous post announcing the open-source of its MLX implementation (https://lnkd.in/dNtdfyzW). The Roofline model helps explain the performance differences by analyzing the computational and memory bandwidth limits of the hardware. I hope this tutorial shares my understanding about performance optimization techniques for Apple silicon GPUs and showcases the outstanding performance of MLX as well as the benefits it brings to applications like Recurrent Drafting. Your feedback is welcome!

119

5 Comments

My Cheeky Bot comes with 32 themes for pro subscribers. Effortlessly change the look and feel of your Cheeky system :) #saas #development

1

Finally hit a major milestone for my ChatGPT Telegram bot GPToid: https://t.me/GPToid_Bot 🇬🇧 Added the English language (previously it was only in Russian). ⭐Added Telegram Stars payments so the product bundles can be purchased from anywhere (the stars can be bought in a local currency and can be used as an intermediate one). There are a lot of AI tools nowadays. What makes GPToid special? 1️⃣ Free tier includes 10 requests to GPT-4o mini daily and 3 DALL-E 3 images weekly (the images are pretty expensive by the way). 2⃣ New users receive a welcome package. Invite others and get rewarded! 3⃣ The bot includes built-in roles and allows configuring a custom prompt. 4⃣ ChatGPT responses are rendered beautifully, especially if you ask about code generation. 5⃣ Quick updates. If OpenAI rolls out a new model, it is added to the bot in a matter of days. 5⃣ Lifetime product bundles. If you buy a package, it won't expire. It's yours forever; use it as you like. You are welcome! Give it a try: https://t.me/GPToid_Bot P.S.: Also, it's open source as all my projects: https://lnkd.in/df5gMNkH #openai #telegram #bots #gptoid #chatgpt

4