About
Activity
2K followers
-
Alexandr Plashchinsky reposted thisAlexandr Plashchinsky reposted thisToday we are excited to announce, in partnership with the Google Kubernetes Engine (GKE) team at Google Cloud, a major milestone in Ray Serve LLM’s throughput and latency characteristics: Ray Serve LLM now matches high performance, rust-based routing frameworks such as vllm-router in benchmarks across a variety of workloads and deployment patterns. In our new blog, we cover three major optimizations to the Ray Serve LLM + vLLM stack that made this possible: direct streaming, a new vLLM Ray executor backend, and HAProxy integration. As a result, we see up to 4.4x higher request throughput than previous versions on prefill-heavy workloads, and up to 24x higher request throughput on decode-heavy workloads. Ray is a popular choice for complex, Python-native distributed computing batch inference pipelines with heterogeneous hardware. And now, we believe that Ray’s powerful primitives for fault tolerance, observability, flexibility across Kubernetes and VMs will enable the next generation of optimizations as LLM inference deployments become increasingly complex. Thanks to Spencer Peterson, Andrew Sy Kim, Kourosh Hakhamaneshi, Jeffrey (Yu-Che) Wang, Richard Liaw, Akshay Malik, Abrar Sheikh, and Alex Yang whose contributions to Ray Serve and Ray Serve LLM made this possible. A special thanks to the vllm-router (vLLM) and SGLang Model Gateway (SGLang) teams for great engineering on their respective projects. Read the full writeup here: https://lnkd.in/guHrz_FA
-
Alexandr Plashchinsky shared thisI am excited to announce that I have been accepted into Johns Hopkins Whiting School of Engineering to pursue my masters degree, where I will be studying Robotics, Machine Learning, & human-machine interaction!
-
Alexandr Plashchinsky shared thisQuick update on Project Cerebrus, my EEG-controlled DIY humanoid robot. Last week, I completed the v1 tech demo: live EEG/biopotential signals → Python pipeline → deep learning inference → Socket.IO relay → physical humanoid robot movement. The v1 demo worked, but it was very jittery. The full system was functional, but the live classification behavior was not stable enough for a polished control interface. Over the past week, I made a lot of changes to improve the model training & production pipeline: - Changed the physical command interface to more detectable actions for my minimal hardware setup: left head tilt, right head tilt, jaw clench, eyebrow raise, and rest - Improved electrode placement by placing the electrodes on both sides of my hairline for a more repeatable frontal lobe placement - Increased the hardware sampling rate to 1000 Hz, from the initial 100 Hz setting - Recorded 90 seconds of training data per class instead of 60 - Simplified preprocessing down to reference-electrode subtraction only - Fixed a major training/production mismatch: both training and live inference now preprocess each 1000-sample window the same way - Changed training window stride to 250 samples instead of 1 to reduce redundant training examples - Added softmax confidence scoring in the live pipeline so the robot can filter out low-confidence predictions before executing commands The result: v2 works significantly better. The robot’s movements are now much more consistently responsive to the live classifier. It is still not perfect, left/right head tilt can still be noisy, but jaw clench and eyebrow raise are being detected much more reliably, and the whole system feels much closer to an actual usable biological-signal interface. The model results were also more interesting this time around. In v1, the flattened MLP outperformed the Transformer encoder, which suggested that the data/preprocessing setup was probably not giving the sequence model much meaningful temporal structure to learn from. In v2, after increasing the sampling rate to 1000 Hz and cleaning up the pipeline, the Transformer encoder started significantly outperforming the flattened MLP. That was encouraging because it suggests the model is now seeing more useful temporal structure in the signal. The best-performing model overall is still the subtractive convolutional architecture, which is designed to learn from channel contrast/difference patterns. check out the updated code here: https://lnkd.in/gK6VR4Tm
-
Alexandr Plashchinsky shared thisI just completed the v1 tech demo of Project Cerebrus, something I’ve been working on for the past few months: an EEG-controlled DIY humanoid robot. The system takes live EEG/biopotential signals from electrodes, streams them into a Python pipeline through Lab Streaming Layer, preprocesses the signal, runs inference through a deep learning model, sends the predicted command through a Socket.IO server deployed on AWS, and then triggers movement sequences on a physical humanoid robot. The current model is a subtractive convolutional neural network trained to classify command-like signals such as left fist clench, right fist clench, jaw clench, and rest. In offline validation, the best subtractive convolutional model reached roughly 85% accuracy, outperforming the other architectures I tested. The real-time control, however, is still very imperfect. The robot can walk forward and turn left/right based on classifier outputs, but the live behavior is jittery and not yet stable enough for a polished control interface. A big reason is that the current setup is bottlenecked by the EEG collection process itself: low channel count, adhesive electrodes, subtle placement inconsistencies, noisy biological signals, and session-to-session variation. Small changes in electrode placement can noticeably change the classification behavior. So while the v1 demo works somewhat, it also made the core limitation very clear: the hardware/software integration is functional, but the classifier & hardware reliability is not where it needs to be yet. This v1 was mainly about proving the full end-to-end system: EEG signal → preprocessing → deep learning inference → websocket relay → robot command execution Now that the full pipeline works, I’m shifting into the research side of the project: optimizing deep learning architectures for EEG classification in robotic command interfaces using higher-quality public EEG datasets. Check out my code here: https://lnkd.in/gK6VR4Tm
-
Alexandr Plashchinsky reposted thisAlexandr Plashchinsky reposted thisRay Summit is back. San Francisco, Marriott Marquis, August 24–26. Three days on the four frontiers shaping AI right now: → Foundation Model Training → Multimodal Data Curation → Physical AI → LLM Reinforcement Learning Keynotes confirmed from Felix Heide (Torc Robotics), Ioannis Antonoglou (Reflection), and Kevin Peterson (Bedrock Robotics), with more to come. Three ways to be part of it: 🎟️ Register at early bird pricing before rates double 🎤 Submit a talk – CFP is open to teams running Ray in production 🤝 Sponsor the summit and reach the builders shaping what's next in AI Full details: https://lnkd.in/ggnZex9v
-
Alexandr Plashchinsky shared thisOver the past few months, I’ve been working on a research project I’m calling Project Cerebrus. The goal is to build a humanoid robot that can be controlled through Electroencephalography (EEG)-based signals. So far, I’ve bought the robotic hardware, assembled the humanoid, mapped out servo positions for walking and turning, and built the first version of the robot’s movement-control logic. Recently, I’ve been studying neuroscience and EEG signaling to understand how brain activity can be measured from the scalp and converted into usable control signals. One thing I’ve found especially interesting is how neural communication loosely mirrors ideas from deep learning architectures. A neuron receives excitatory and inhibitory postsynaptic potentials, sums those inputs, and only fires if a threshold is reached. Conceptually, that thresholding behavior feels similar to an activation function like ReLU in a neural network: information is only propagated downstream when certain conditions are met. I’ve now assembled the EEG sensor setup and figured out a simple pipeline to stream live EEG data into Python, where the next steps are preprocessing, denoising, and building a small transformer model to classify fixed windows of brain data into robot commands.
-
Alexandr Plashchinsky shared thisAs promised, I have just posted a nearly 4-hour-long YouTube video where I explain & code, end-to-end, my own custom PyTorch-style deep learning optimization engine in c++ :) Here is the full video: https://lnkd.in/gfG99sgC
-
Alexandr Plashchinsky shared thisThis weekend, I finished building my own custom implementation of a deep learning optimization engine in C++, from scratch. I have been working on this project over the past 2 weeks, with the intention of improving my C++ & ML skills. I am planning on recording a video soon where I break down the full theory & implementation, as well as do a full coding tutorial, where I code the engine end-to-end, explaining each line along the way :). For now, here is the public github (the README.md & MNIST data downloader script are fully vibe coded tho 💀) : https://lnkd.in/g3KaCgnb
-
-
Alexandr Plashchinsky reacted on thisAlexandr Plashchinsky reacted on thisMeet our new Robotics Community Lead: Dhruv Diddi Dhruv is an ML engineer, founder, and Meta Llama Impact Grant recipient who's been quietly building something remarkable — right here at Frontier Tower. Through Solo Tech, he's advancing Physical AI Inference: offline, multilingual AI tools designed to reach the people most technology leaves behind, across agriculture, education, and health in underserved rural communities. That's the kind of work that defines the frontier. We couldn't be more excited to have him stepping into this role, and we know he's going to bring that same energy and vision to the Robotics community here. Explore his work at getsolo.tech and connect with him on LinkedIn. Welcome to the lead team, Dhruv.
-
Alexandr Plashchinsky reacted on thisAlexandr Plashchinsky reacted on thisI got to present our team's work on how we scale vision AI inference on satellite imagery to continental scale using Ray on Anyscale. A big thanks to the Anyscale team for hosting us and for the invaluable technical guidance they gave us throughout. https://lnkd.in/e238HpYd Milos Colic Vuong Nguyen Pritimoy Podder Pablo Hidalgo Ryan Bashir Ali Sezer Alexandr PlashchinskyHow Adyen trains a Transaction Foundation Model (TFM) on 51 trillion tokens and other stories on scaling AI with Ray from Xoople, Criteo, and BMW | AnyscaleHow Adyen trains a Transaction Foundation Model (TFM) on 51 trillion tokens and other stories on scaling AI with Ray from Xoople, Criteo, and BMW | Anyscale
-
Alexandr Plashchinsky reacted on thisAlexandr Plashchinsky reacted on thisI’m thrilled to share a few major life and career updates! 🎓 Ph.D. & Best Dissertation Award: I have officially graduated with my Ph.D. from UCLA. I am incredibly honored to share that my dissertation, "Building World Models with Spatial Intelligence," was awarded the 2026 UCLA Distinguished Ph.D. Dissertation Research Award (sole 🏅 Gold recipient of this highest honor). None of this would have been possible without the invaluable guidance of my Ph.D. advisor, Prof. Achuta Kadambi. I also want to extend my deepest gratitude to my Ph.D. committee members, Prof. Bolei Zhou, Prof. Aydogan Ozcan, and Prof. Jonathan Kao, for their unwavering support and feedback. Furthermore, it is a great privilege to thank Prof. Leonidas Guibas from Stanford University and Prof. Atlas Wang from The University of Texas at Austin for their instrumental collaboration, mentorship, and guidance throughout my doctoral journey. 🚀 Next Chapter: I’m excited to announce that I have joined Google as a Research Scientist! I will be focusing on bringing spatial intelligence into the next generation of real-world applications under the leadership of Prof. Cengiz Oztireli. I’m excited to be back in the Bay Area as I continue to push the boundaries of world models, spatial intelligence, and spatial computing. I am always eager to exchange ideas and explore new collaborations, so if you are in the area or share similar research interests, please feel free to reach out and connect!
-
Alexandr Plashchinsky liked thisAlexandr Plashchinsky liked thisThrilled to share that Timothy Seah has joined Reflection AI as a Member of Technical Staff building pretraining infrastructure. Reflection is one of the most exciting AI companies right now, building open-weight frontier models with a $25B valuation and a team that genuinely believes in getting superintelligence right. This was a role that demanded someone with rare depth. Matching engineers to the right frontier AI opportunity is what I love doing. If you or someone you know is working on distributed training, ML systems, or pre-training infrastructure and wants to work on something that matters, I would love to connect. #AIRecruitment #ReflectionAI #MLSystems #PretrainingInfrastructure
-
Alexandr Plashchinsky reacted on thisAlexandr Plashchinsky reacted on thisImagine math is like building the world's biggest LEGO city. 🏗️ Each of these people figured out an important building block that made math much bigger and stronger. ⚙️ Archimedes 🌊 He loved figuring out how things move, float, and balance. 🚢 If you wanted to know why a boat floats or how to lift something really heavy with a lever, he'd help. 👧 Kid version: "The scientist who figured out the secrets of pushing, pulling, and floating." 📐 Euclid 📏 He organized geometry into simple rules. 🧩 He taught that if you know a few basic rules, you can solve thousands of problems. 👧 Kid version: "The teacher who wrote the rulebook for shapes." ➕➖ Al-Khwarizmi 🔤 He helped invent algebra, where letters can stand for numbers. 💻 The word algorithm (what computers use to solve problems) comes from his name! 👧 Kid version: "The puzzle master who taught us how to solve number mysteries." 🍎 Isaac Newton 📈 He created calculus, a way to study things that are always changing. 🚀 It helps us understand planets, rockets, cars, and even roller coasters. 👧 Kid version: "The inventor of math for things that move." 🔢 Carl Friedrich Gauss 🔍 He was incredibly good at finding patterns in numbers. 👑 Many people call him the "Prince of Mathematics." 👧 Kid version: "The detective who found hidden patterns in numbers." 🌐 Leonhard Euler ✍️ He wrote so much math that we still use his symbols today. 🕸️ He connected many different parts of mathematics together. 👧 Kid version: "The great organizer who connected all the math ideas." 🎮 Think of it like building a video game: 🏗️ Archimedes built the physics engine. 📐 Euclid designed the map and shapes. 🧩 Al-Khwarizmi wrote the puzzle-solving instructions. 🚀 Newton made everything move realistically. 🔍 Gauss found the hidden patterns and shortcuts. 🌐 Euler connected everything so it worked together. 💡 One small note The image says these are the "Six Fathers of Mathematics," but that's a simplification. There isn't one official list of "fathers of mathematics." These six are among the most influential mathematicians in history, but many others—such as Pythagoras, René Descartes, and Srinivasa Ramanujan—also made enormous contributions. 🌍 It's better to think of these six as six giants whose ideas helped build the mathematical world we use today. ✨ Mathematics isn't just about numbers—it's the language humanity uses to understand, describe, and build the universe
-
Alexandr Plashchinsky reacted on thisAlexandr Plashchinsky reacted on this600+ users. 6 months. Exactly 1 day with zero signups. That last metric is the one I'm most proud of. Since launching RoboSpace.app, June 29th was the only day a new builder didn't show up. Every single other day, someone arrived ready to build. Now, we’re taking the next step: partnering directly with university research labs and robotics teams. If you build robot simulations, I want to hear from you: 1. What does your current workflow look like? 2. Where do you lose the most time iterating?
Experience & Education
-
Anyscale
View Alexandr’s full experience
See their title, tenure and more.
or
Publications
-
Invertible Memory Flow Networks
Long sequence neural memory remains a challenging problem. RNNs and their variants suffer from vanishing gradients, and Transformers suffer from quadratic scaling. Furthermore, compressing long sequences into a finite fixed representation remains an intractable problem due to the difficult optimization landscape. Invertible Memory Flow Networks (IMFN) make long sequence compression tractable through factorization: instead of learning end-to-end compression, we decompose the problem into…
Long sequence neural memory remains a challenging problem. RNNs and their variants suffer from vanishing gradients, and Transformers suffer from quadratic scaling. Furthermore, compressing long sequences into a finite fixed representation remains an intractable problem due to the difficult optimization landscape. Invertible Memory Flow Networks (IMFN) make long sequence compression tractable through factorization: instead of learning end-to-end compression, we decompose the problem into pairwise merges using a binary tree of "sweeper" modules. Rather than learning to compress long sequences, each sweeper learns a much simpler 2-to-1 compression task, achieving O(log N) depth with sublinear error accumulation in sequence length. For online inference, we distilled into a constant-cost recurrent student achieving O(1) sequential steps. Empirical results validate IMFN on long MNIST sequences and UCF-101 videos, demonstrating compression of high-dimensional data over long sequences.
Other authors -
-
Parent-Guided Semantic Reward Model (PGSRM): Embedding-Based Reward Functions for Reinforcement Learning of Transformer Language Models
See publicationWe introduce the Parent-Guided Semantic Reward Model (PGSRM), a lightweight reward framework for reinforcement learning (RL) of transformer language models. PGSRM replaces binary correctness signals, human preference data, and trained reward models with a simple signal: cosine similarity between a parent model's reference output embedding and a child model's generated output for the same input. This yields a dense, semantically meaningful reward with no human annotation or additional model…
We introduce the Parent-Guided Semantic Reward Model (PGSRM), a lightweight reward framework for reinforcement learning (RL) of transformer language models. PGSRM replaces binary correctness signals, human preference data, and trained reward models with a simple signal: cosine similarity between a parent model's reference output embedding and a child model's generated output for the same input. This yields a dense, semantically meaningful reward with no human annotation or additional model training. We apply PGSRM on five language tasks and find that it produces smoother reward improvement and more stable PPO dynamics than a binary reward baseline, suggesting that embedding-based semantic rewards are a practical alternative to RLHF-style reward modeling for parent-guided alignment in smaller transformer models.
Projects
-
Project Cerebrus: EEG-controlled Humanoid Robot
-
Project Cerebrus is an EEG-controlled robotics project that connects a PLUX NeuroBIT EEG hardware kit to a DIY humanoid robot. The system streams live data from the EEG hardware, classifies the signal into movement commands using subtractive convolutional layers, sends those commands through a websocket relay server, and executes movement sequences on the robot.
-
C++ Deep Learning Optimization Engine
-
A small C++ project, for a minimal deep learning optimization engine from scratch. The project includes a minimal Tensor class with automatic differentiation, matrix/vector operations, activation functions, softmax, cross-entropy-style loss construction, and SGD parameter updates.
-
TinyGPU
-
TinyGPU is a compact SystemVerilog matrix-vector accelerator built around a clean streaming architecture: a fetch_engine issues memory reads and pushes data through a skid-buffered valid/ready stream into a matrix_core that loads weights, accumulates results, and flushes an output vector. The top module (gpu_top) ties host control, memory interface, and result streaming together, exposing a simple programming model (set base addresses → RUN) with a unified busy status.
-
Project Benji: Autonomous Navigational/Conversational Robot
-
Autonomous patrol + conversational robot, using a self-hosted GPT language model, Whisper STT, and ElevenLabs TTS. Implements SLAM-style navigation with ultrasonic sensors, multi-user identity memory, agentic movement/vision commands, and a custom Q-table RL algorithm with sentiment analysis to adapt Benji’s behavior per user.
-
Learning Rate - AutoML Platform
-
Learning Rate is an AutoML platform to simplify machine learning for non-experts, allowing users to upload their own data and quickly generate a vast array of ML models.
Tech Stack: Next.js for front-end development, Flask & Sci-Kit Learn for backend and machine learning, and GCP as data lake for model & data storage. -
Reinforcement Learning Algorithms - DQN, PPO, Q-Learning
-
Utilized DQN algorithm for learning Q-values from raw state observations, achieving a trained model capable of navigating the Frozen Lake environment.
Implemented PPO algorithm to optimize policy updates, resulting in a trained model proficient in controlling the Bipedal Walker environment.
Used tabular Q-learning for optimal action selection, training a model capable of playing blackjack. -
Blocky - NLP + Blockchain Data Marketplace
-
Blocky uses a fine-tuned bloom-3b model, as well as a 300-dimensional vector space model to match data seekers with providers, via a vector database architecture.
Users configure AI agents, powered by the Gemini-Pro LM, to negotiate data transactions via the ETH main net.
Tech Stack: Next.js, Firebase, Flask, Bloom-3b LM, FastText VS Model, Gemini-Pro LM, Solidity. -
Sentinel - Social Media Fraud Detection
-
Developed Sentinel, an LLM to detect fraudulent activities for platforms/mediums such as YouTube, Twitter, Instagram, and SMS.
Integrated Apache Spark for data processing, transformer architecture + BERT model for NLP, and Next.js for frontend design. -
Prima - Fashion Trend Analysis Tool
-
Developed a Python-based fashion trend analysis tool, incorporating PostgreSQL for data management, Node.js for server-side operations, and Next.js for user interface design.
Utilized Apache Airflow automated web scraping process, and SQL data wrangling to identify and analyze fashion trends. -
Theseus - nanoGPT
-
Engineered a transformer-based language model to generate Greek mythology narratives, utilizing Python + PyTorch and a custom dataset.
Configured model with multi-head attention, layer normalization, and feed-forward networks; optimized for CPU with a modified architecture and hyperparameters. -
RePrice - Web Scraping Pricing/Ratings Comparison Tool
-
Developed a web app with Typescript and Python to scrape online data for monthly pricing and ratings comparisons.
Uses AI algorithm to estimate monthly pricing plans of top competitors for any inputted product category.
Stores product information in postreSQL database. -
WebAnalytics - ETL data pipeline + custom KPI dashboard
-
Developed an automated web analytics tool leveraging GA4, GCP, and AppScript to streamline the collection and parsing of web analytics data.
Features a custom dashboard for intuitive display and analysis of key web performance metrics, enhancing data-driven decision-making. -
MarketDriven - Strategic Financial Planning Model
-
Developed a financial forecasting model using Excel to predict income statements, balance sheets, and cash flow statements up to seven years ahead.
Incorporates user-inputted COGS, OPEX, and their projected growth, alongside marketing campaign data, estimated conversion rates, and churn rates.
Outputs comprehensive financial forecasts to aid in strategic planning and decision-making.
Test Scores
-
SAT
Score: 1540
Score Breakdown:
Math: 800, Reading: 740 -
SAT Math Level II Subject Test
Score: 800
Languages
-
Russian
Native or bilingual proficiency
-
English
Native or bilingual proficiency
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content