Paul Iusztin

Stealth AI Startup

Founding Al Engineer

Decoding ML

Content to help you design, code, and deploy production-grade AI, GenAI and information retrieval systems.

Founder

Remote Work

💼 Senior machine learning engineer, contractor, consultant, and freelancer with 𝟔+ 𝐲𝐞𝐚𝐫𝐬 𝐨𝐟 𝐞𝐱𝐩𝐞𝐫𝐢𝐞𝐧𝐜𝐞 offering scalable and modular machine learning solutions for businesses worldwide.

➞ I can design, implement, train, evaluate, and operate end-to-end ML pipelines for your business. I can scale up your ML systems by leveraging the latest MLOps technologies.

➞ I have spent 7+ years writing code in Python, but I also enjoy using Node.js for web development and C/C++ for optimizing various applications.

➞ My go-to machine learning frameworks are PyTorch, Keras, and Sklearn. Also, my favorite MLOps tech stack is AWS, W&B/MLFlow, and DVC.

🔥The world of AI is constantly changing, new technologies are born almost daily, and therefore, I am always open to new challenges.

💬 𝐂𝐨𝐧𝐭𝐚𝐜𝐭 𝐦𝐞, and let's see how I can help your business.

Senior Machine Learning Engineer • Freelancer • Contractor

Metaphysic.ai

🟡 Computer vision & deep learning workflow engine: orchestrate & execute

The client, a leading GenAI platform, faced difficulties scaling and breaking down its complex ML jobs. Their manual processes and CLI-based tooling added significant friction in serving their clients. Thus, the client required a more intuitive and scalable tool that could be used across the company to train and serve its ML models.

→ 𝗥𝗲𝗳𝗮𝗰𝘁𝗼𝗿𝗲𝗱 the ML worker architecture to follow DDD principles, making it more 𝗿𝗲𝗮𝗱𝗮𝗯𝗹𝗲, 𝗺𝗼𝗱𝘂𝗹𝗮𝗿, and 𝗲𝘅𝘁𝗲𝗻𝘀𝗶𝗯𝗹𝗲.

→ 𝗗𝗲𝗰𝗿𝗲𝗮𝘀𝗲𝗱 the 𝗹𝗮𝘁𝗲𝗻𝗰𝘆 of the data curation DL models 𝗯𝘆 𝟴𝟮.𝟳%

→ 𝗘𝘅𝘁𝗲𝗻𝗱𝗲𝗱 their clustering algorithm to accurately run on 𝗱𝗮𝘁𝗮𝘀𝗲𝘁𝘀 >𝟭𝟬𝗸 𝗶𝗺𝗮𝗴𝗲𝘀 and 𝗮𝘂𝘁𝗼𝗺𝗮𝘁𝗶𝗰𝗮𝗹𝗹𝘆 𝗱𝗲𝘁𝗲𝗰𝘁 the number of 𝗰𝗹𝘂𝘀𝘁𝗲𝗿𝘀 and 𝗼𝘂𝘁𝗹𝗶𝗲𝗿𝘀.

Senior Machine Learning Engineer

⚫️ Scalable ML batch pipeline for continuous training & inference

The client, a leading GenAI platform, struggled with manual training and calling their ML models. As a tedious and unscalable process, they needed to automate their ML processes completely and integrate them with their current web platform.

→ 𝗗𝗲𝘀𝗶𝗴𝗻𝗲𝗱 and 𝗶𝗺𝗽𝗹𝗲𝗺𝗲𝗻𝘁𝗲𝗱 a 𝗺𝗼𝗱𝘂𝗹𝗮𝗿 𝗣𝘆𝘁𝗵𝗼𝗻 𝗽𝗮𝗰𝗸𝗮𝗴𝗲 that orchestrates the ML steps into 3 fully automated batch pipelines: feature, training, and inference, 𝗲𝗹𝗶𝗺𝗶𝗻𝗮𝘁𝗶𝗻𝗴 𝗮 𝘁𝗲𝗱𝗶𝗼𝘂𝘀 𝟮-𝗵𝗼𝘂𝗿 𝗺𝗮𝗻𝘂𝗮𝗹 𝗽𝗿𝗼𝗰𝗲𝘀𝘀/𝘃𝗶𝗱𝗲𝗼.

→ 𝗥𝗲𝗱𝘂𝗰𝗲𝗱 the 𝗽𝗿𝗼𝗰𝗲𝘀𝘀𝗶𝗻𝗴 𝘁𝗶𝗺𝗲 of the batch ML pipeline 𝗯𝘆 𝟲𝟮.𝟱% while preserving the same accuracy and quality.

→ 𝗜𝗺𝗽𝗹𝗲𝗺𝗲𝗻𝘁𝗲𝗱 𝗠𝗟𝗢𝗽𝘀 𝗯𝗲𝘀𝘁 𝗽𝗿𝗮𝗰𝘁𝗶𝗰𝗲𝘀 such as IaC, CI/CD, monitoring, experiment tracking, and model registries, making the system reproducible, testable and trackable.

→ 𝗗𝗲𝗽𝗹𝗼𝘆𝗲𝗱 a scalable & cost-effective 𝗮𝘀𝘆𝗻𝗰 𝗯𝗮𝘁𝗰𝗵 𝗮𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲 on top of AWS ECS & SQS that scales in and out based on the number of jobs, 𝗿𝗲𝗱𝘂𝗰𝗶𝗻𝗴 their 𝗔𝗪𝗦 𝗰𝗼𝘀𝘁𝘀 𝗯𝘆 𝟱𝟮%.

⚫️ Big data ingestion & archiving pipeline

The client wanted to expand its platform features by securely ingesting, preprocessing, and storing TB of data into their platform.

→ 𝐂𝐨𝐧𝐭𝐫𝐢𝐛𝐮𝐭𝐞𝐝 to the 𝐝𝐞𝐬𝐢𝐠𝐧 and 𝐢𝐦𝐩𝐥𝐞𝐦𝐞𝐧𝐭𝐚𝐭𝐢𝐨𝐧 of a 𝐝𝐚𝐭𝐚 𝐢𝐧𝐠𝐞𝐬𝐭𝐢𝐨𝐧 𝐩𝐢𝐩𝐞𝐥𝐢𝐧𝐞 that supports 𝐢𝐧𝐠𝐞𝐬𝐭𝐢𝐧𝐠 >𝟓 𝐓𝐁 of data in 𝐨𝐧𝐞 𝐠𝐨 in a reliable, fast, and secure way.

→ 𝗜𝗺𝗽𝗹𝗲𝗺𝗲𝗻𝘁𝗲𝗱 a scalable 𝗱𝗮𝘁𝗮 𝗽𝗿𝗲𝗽𝗿𝗼𝗰𝗲𝘀𝘀𝗶𝗻𝗴 𝗽𝗶𝗽𝗲𝗹𝗶𝗻𝗲 that supports multiple media types (images, videos, audio, 3D objects), archiving the original files on an affordable storage solution, 𝗿𝗲𝗱𝘂𝗰𝗶𝗻𝗴 𝗰𝗼𝘀𝘁𝘀 𝗯𝘆 𝘂𝗽 𝘁𝗼 𝟴𝟭%.

Towards Data Science

⚫️ ML and MLOps engineering blog

Helping machine learning engineers level up through hands-on practical articles and tutorials on designing and productionizing ML systems.

➞ Published 𝐓𝐡𝐞 𝐅𝐮𝐥𝐥 𝐒𝐭𝐚𝐜𝐤 𝟕-𝐒𝐭𝐞𝐩𝐬 𝐌𝐋𝐎𝐩𝐬 𝐅𝐫𝐚𝐦𝐞𝐰𝐨𝐫𝐤 with 90k+ views on Medium and 600+ stars on GitHub. An MLE & MLOps course that teaches you how to design, build, and deploy an end-to-end ML batch system: https://github.com/iusztinpaul/energy-forecasting

Machine Learning Engineer / Technical Writer

CoreAI

⚫️ Real-time social media recommender system using machine learning & Kafka

The client, a social media start-up from Israel, needed a real-time dashboard and recommender system on top of their Kafka infrastructure using modern ML and MLOps solutions.

➞ 𝐋𝐞𝐝 the architecture system design of a modular real-time social media recommender system built on top of the client's Kafka nervous system that leverages streaming and ML technologies such as Kafka Streams, Kotlin, PyTorch, Python, and Docker.

➞ Quickly 𝐩𝐥𝐚𝐧𝐧𝐞𝐝 and 𝐝𝐞𝐯𝐞𝐥𝐨𝐩𝐞𝐝 a scalable recommender system based on heuristics, picked through A/B testing, using Kafka Streams and Kotlin. It served as a vital entry point and baseline for future algorithms.

➞ 𝐃𝐞𝐬𝐢𝐠𝐧𝐞𝐝 and 𝐢𝐦𝐩𝐥𝐞𝐦𝐞𝐧𝐭𝐞𝐝 a dashboard that computes KPIs in 𝐫𝐞𝐚𝐥-𝐭𝐢𝐦𝐞 𝐟𝐨𝐫 𝟏𝟎𝟎𝐤+ 𝐮𝐬𝐞𝐫𝐬 on the client's Kafka infrastructure using ksqlDB, SQL, Elasticsearch, and Apache Superset.

Machine Learning Tech Lead

⚫️ coreControl - end-to-end MLOps infrastructure

The client, a machine learning service provider, required an end-to-end ML/MLOps infrastructure to scale their operations and save development resources across multiple AI projects.

➞ 𝐋𝐞𝐝 the 𝐝𝐞𝐬𝐢𝐠𝐧 and 𝐢𝐦𝐩𝐥𝐞𝐦𝐞𝐧𝐭𝐚𝐭𝐢𝐨𝐧 of coreControl by integrating the software with the client's products and collaborating with other ML engineers, developers, and graphic designers.

➞ 𝐑𝐞𝐝𝐮𝐜𝐞𝐝 the 𝐝𝐞𝐯𝐞𝐥𝐨𝐩𝐦𝐞𝐧𝐭 𝐢𝐦𝐞 across the client's machine learning projects 𝐛𝐲 𝟑𝟎% using coreControl.

➞ 𝐈𝐦𝐩𝐥𝐞𝐦𝐞𝐧𝐭𝐞𝐝 the data backbone of coreControl that syncs the data between multiple MLOps tools and exposes them to a UI through a RESTful API using Python, FastAPI, PostgreSQL, Docker, and AWS.

➞ 𝐃𝐞𝐯𝐞𝐥𝐨𝐩𝐞𝐝 a Python package that wraps various MLOps tools that standardizes the data scientists' work and centralizes all the ML artifacts to easily be analyzed in a single place using Python, DVC, and ClearML.

⚫️ Prototyping AI solutions for Demand Forecasting, Churn Prediction, and Customer Segmentation

The client, a machine learning start-up, wanted to expand their services portfolio by prototyping and gaining knowledge in various AI fields that leverage structured retail data.

➞ 𝐋𝐞𝐝 𝐚 𝐭𝐞𝐚𝐦 𝐨𝐟 𝟐 to develop an ML clustering solution that aggregates data from multiple sources and finds customer segments and insights for new potential clients.

➞ 𝐈𝐦𝐩𝐥𝐞𝐦𝐞𝐧𝐭𝐞𝐝 a prototype for churn and demand forecasting used as a proof of concept to close deals with various clients.

➞ 𝐒𝐚𝐯𝐞𝐝 𝟒𝟎% of 𝐞𝐧𝐠𝐢𝐧𝐞𝐞𝐫𝐢𝐧𝐠 𝐫𝐞𝐬𝐨𝐮𝐫𝐜𝐞𝐬 by building a reusable Python ML framework that can quickly be tailored to custom problems based on structured data.

Politehnica University of Timisoara

⚫️ Teaching the foundations of 𝐚𝐫𝐭𝐢𝐟𝐢𝐜𝐢𝐚𝐥 𝐢𝐧𝐭𝐞𝐥𝐥𝐢𝐠𝐞𝐧𝐜𝐞 laboratory.

➞ Supervised and unsupervised learning
➞ Linear regression, SVMs, Decision trees
➞ Clustering, Dimensionality reduction
➞ Dense neural networks
➞ Convolutional neural networks
➞ Recurrent neural networks
➞ Transformers

Teaching Assistant

Everseen

🟡 Autonomous supply chain system using computer vision & deep learning

The client, a well-known player in the retail sector, needed to expand its deep learning and computer vision solutions by researching and building autonomous systems for its operations in the retail industry.

➞ 𝐓𝐫𝐚𝐢𝐧𝐞𝐝 a Mask-RCNN model for 2D object detection and segmentation on 100k proprietary retail images, with an 𝐦𝐀𝐏 𝐨𝐟 𝟖𝟓%.

➞ 𝐒𝐩𝐞𝐞𝐝 𝐮𝐩 the 𝐢𝐧𝐟𝐞𝐫𝐞𝐧𝐜𝐞 𝐭𝐢𝐦𝐞 of the Mask-RCNN model 𝐛𝐲 𝟓𝟎% by levering Docker, ONNX, and TensorRT to containerize and accelerate the models.

➞ 𝐓𝐫𝐚𝐢𝐧𝐞𝐝 a scalable classifier using an image retrieval model that can dynamically handle new classes without additional training.

➞ 𝐈𝐧𝐜𝐫𝐞𝐚𝐬𝐞𝐝 the 𝐫𝐞𝐩𝐫𝐨𝐝𝐮𝐜𝐢𝐛𝐢𝐥𝐢𝐭𝐲 of the 𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐩𝐫𝐨𝐜𝐞𝐬𝐬 by implementing MLOps solutions, such as DVC & MLFlow, that included data & model versioning, registries, and experiment tracking features.

Continental

🟡 Autonomous driving system using computer vision & deep learning

The client, a well-known player in the automotive sector, needed to create an autonomous driving car prototype that could accurately detect traffic participants by leveraging DL and CV solutions.

➞ 𝐋𝐞𝐝 the 𝐝𝐞𝐯𝐞𝐥𝐨𝐩𝐦𝐞𝐧𝐭 of a prototype for real-time traffic participants detection by deploying a CenterNet model for 3D object detection on a test car using an Nvidia Drive AGX C++ platform.

➞ 𝐈𝐦𝐩𝐫𝐨𝐯𝐞𝐝 the reproducibility of the experimentation process by implementing MLOps solutions (DVC) that version unstructured big data.

➞ 𝐈𝐦𝐩𝐥𝐞𝐦𝐞𝐧𝐭𝐞𝐝 a replacement to the Hungarian algorithm that 𝐢𝐧𝐜𝐫𝐞𝐚𝐬𝐞𝐝 the 𝐬𝐩𝐞𝐞𝐝 by 𝟑𝟎%, making the system 𝐫𝐞𝐚𝐥-𝐭𝐢𝐦𝐞.

➞ 𝐒𝐚𝐯𝐞𝐝 𝟐 𝐝𝐚𝐲𝐬, in 𝐞𝐯𝐞𝐫𝐲 𝐦𝐨𝐧𝐭𝐡, of 𝐫𝐞𝐩𝐞𝐭𝐢𝐭𝐢𝐯𝐞 work by automating the steps for the creation of the 𝐦𝐨𝐧𝐭𝐡𝐥𝐲 demo.

Mid-Level Machine Learning Researcher

⚫️ Autonomous driving research using computer vision & deep learning

The client, a well-known player in the automotive sector, needed to expand its research into the autonomous driving field using novel deep learning and computer vision methods.

➞ 𝐑𝐞𝐬𝐞𝐚𝐫𝐜𝐡 in 3D object detection, depth estimation & completion, tracking, and sensor fusion algorithms using various sensors such as images, radar & LiDAR point clouds.

➞ 𝐑𝐞𝐝𝐮𝐜𝐞𝐝 𝐛𝐲 𝟒𝟎% the 𝐦𝐞𝐦𝐨𝐫𝐲 𝐮𝐬𝐚𝐠𝐞 of the 3D object detection code by introducing the Dynamic Voxelization algorithm into the preprocessing pipeline.

➞ 𝐓𝐫𝐚𝐢𝐧𝐞𝐝 state-of-the-art models to generate LiDAR-like dense point clouds from sparse Radar point clouds on various autonomous driving datasets, such as KITTI, nuScenes, DDAD, and A2D2.

Junior Machine Learning Researcher

Elementum Technologies

⚫️ Dorel - An Uber for handymen

𝐃𝐨𝐫𝐞𝐥 is an app that connects handypersons with customers. It contains a review and portfolio system where any handyperson can build his profile. Based on that, a user can determine whether he fits the job. The app intends to bring, once again, trust in handypersons.

➞ 𝐏𝐢𝐭𝐜𝐡𝐞𝐝 the 𝐢𝐝𝐞𝐚 to various start-up incubators.

➞ 𝐌𝐚𝐧𝐚𝐠𝐞𝐝 the backend components of the software solution of Dorel using an Agile methodology.

➞ 𝐋𝐞𝐝 the 𝐝𝐞𝐬𝐢𝐠𝐧 and 𝐢𝐦𝐩𝐥𝐞𝐦𝐞𝐧𝐭𝐚𝐭𝐢𝐨𝐧 of the Django RESTful API server and the notifications microservice.

➞ 𝐃𝐞𝐬𝐢𝐠𝐧𝐞𝐝 and 𝐝𝐞𝐯𝐞𝐥𝐨𝐩𝐞𝐝 the Redux layer of the React Native mobile app.

Co-Founder

🤖 Senior Machine Learning Engineer • MLOps • Contractor ~ Helping startups engineer production-ready ML/AI systems.

I am a senior AI engineer and contractor with 8+ years of experience. I design and implement modular, scalable, and production-ready ML systems for startups worldwide.

My central mission is to build an AI academy and data-intensive AI/ML products that serve the world.

Since training my first neural network in 2017,  I have 2 passions that fuel my mission:

→ Designing and implementing production AI/ML systems using software engineering and MLOps best practices.

→ Teaching people about the process.

.

I currently develop vertical AI agent solutions for a US startup. In the past, I built AI, ML and MLOps products for Metaphysic, CoreAI, Everseen, and Continental.

I am the Founder of Decoding ML, a channel for battle-tested content on learning how to design, code, and deploy production-grade AI systems.

As a content creator, I co-authored the LLM Engineer’s Handbook, an Amazon bestseller and industry standard on LLM, RAG, and LLMOps solutions, which sold over 15k+ copies.

.

If you want to learn how to build end-to-end production-ready AI systems with software engineering and MLOps best practices, subscribe to my Substack:

🔗 https://decodingml.substack.com

.

💬 If you need AI solutions for your business, let’s discuss!

🌎 Only open to full remote positions as a contractor.

.

Contact:

✉️ Email: p.b.iusztin@gmail.com

💻 Decoding ML: https://decodingml.substack.com

🟡 Scalable ML batch infrastructure for continuous training & inference

The client, a generative AI leader, struggled with manual training and calling their ML models. As a tedious and unscalable process, they needed to automate their ML processes completely and integrate them with their current web platform.

➞ 𝐃𝐞𝐬𝐢𝐠𝐧𝐞𝐝 and 𝐢𝐦𝐩𝐥𝐞𝐦𝐞𝐧𝐭𝐞𝐝 a modular 𝐏𝐲𝐭𝐡𝐨𝐧 𝐩𝐚𝐜𝐤𝐚𝐠𝐞 that composes the ML model required steps into 3 main batch pipelines that can easily be automated: feature, training, and inference, 𝐞𝐥𝐢𝐦𝐢𝐧𝐚𝐭𝐢𝐧𝐠 a 𝐭𝐞𝐝𝐢𝐨𝐮𝐬 𝟐-𝐡𝐨𝐮𝐫 𝐦𝐚𝐧𝐮𝐚𝐥 𝐩𝐫𝐨𝐜𝐞𝐬𝐬/𝐯𝐢𝐝𝐞𝐨.
➞ 𝐃𝐞𝐩𝐥𝐨𝐲𝐞𝐝 a scalable & cost-effective 𝐛𝐚𝐭𝐜𝐡 𝐚𝐫𝐜𝐡𝐢𝐭𝐞𝐜𝐭𝐮𝐫𝐞 and 𝐂𝐈/𝐂𝐃 𝐩𝐢𝐩𝐞𝐥𝐢𝐧𝐞 on top of AWS ECS & SQS that scales in and out based on the number of jobs, 𝐬𝐚𝐯𝐢𝐧𝐠 𝐭𝐡𝐨𝐮𝐬𝐚𝐧𝐝𝐬 𝐨𝐟 𝐝𝐨𝐥𝐥𝐚𝐫𝐬/𝐦𝐨𝐧𝐭𝐡.
➞ 𝐒𝐚𝐯𝐞𝐝 𝐦𝐨𝐧𝐭𝐡𝐬 𝐨𝐟 𝐟𝐮𝐭𝐮𝐫𝐞 𝐝𝐞𝐯𝐞𝐥𝐨𝐩𝐦𝐞𝐧𝐭 by using good software practices to make the Python module & batch cloud infrastructure flexible and composable.
➞ 𝐁𝐮𝐢𝐥𝐭 a 𝐯𝐨𝐢𝐜𝐞 𝐬𝐲𝐧𝐭𝐡𝐞𝐬𝐢𝐬 𝐏𝐨𝐂 in ~1 week required to 𝐜𝐥𝐨𝐬𝐞 the 𝐝𝐞𝐚𝐥 with one of the 𝐜𝐮𝐫𝐫𝐞𝐧𝐭 𝐜𝐥𝐢𝐞𝐧𝐭𝐬.

⚫️ Big data ingestion & archiving pipeline

The client wanted to expand its platform features by securely and safely ingesting, preprocessing, and storing TB of data into their platform.

➞ 𝐂𝐨𝐧𝐭𝐫𝐢𝐛𝐮𝐭𝐞𝐝 to the 𝐝𝐞𝐬𝐢𝐠𝐧 and 𝐢𝐦𝐩𝐥𝐞𝐦𝐞𝐧𝐭𝐚𝐭𝐢𝐨𝐧 of a 𝐝𝐚𝐭𝐚 𝐢𝐧𝐠𝐞𝐬𝐭𝐢𝐨𝐧 𝐩𝐢𝐩𝐞𝐥𝐢𝐧𝐞 that supports 𝐢𝐧𝐠𝐞𝐬𝐭𝐢𝐧𝐠 >𝟓 𝐓𝐁 of data in 𝐨𝐧𝐞 𝐠𝐨 in a reliable, fast, and secure way.
➞ 𝐈𝐦𝐩𝐥𝐞𝐦𝐞𝐧𝐭𝐞𝐝 a scalable 𝐝𝐚𝐭𝐚 𝐩𝐫𝐞𝐩𝐫𝐨𝐜𝐞𝐬𝐬𝐢𝐧𝐠 𝐩𝐢𝐩𝐞𝐥𝐢𝐧𝐞 that supports multiple media types (images, videos, audio, 3D objects) and creates small preview files that are cheap to transfer, store and visualize. The 𝐨𝐫𝐢𝐠𝐢𝐧𝐚𝐥 𝐟𝐢𝐥𝐞𝐬 are 𝐚𝐫𝐜𝐡𝐢𝐯𝐞𝐝 on an affordable storage solution 𝐫𝐞𝐝𝐮𝐜𝐢𝐧𝐠 𝐜𝐨𝐬𝐭𝐬 𝐛𝐲 𝐮𝐩 𝐭𝐨 𝟖𝟏%.

Everybody talks about shipping AI products, but nobody shows you how.

These 5 books are all you need (and yes... I've read them all):

𝟭/ 𝗗𝗲𝘀𝗶𝗴𝗻𝗶𝗻𝗴 𝗠𝗮𝗰𝗵𝗶𝗻𝗲 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝗦𝘆𝘀𝘁𝗲𝗺𝘀 𝗯𝘆 Chip Huyen 
🔗: https://lnkd.in/dwmcghyA

One of the best books I've read on designing ML systems.

It's beginner-friendly.

And it provides a strong intuition into what an end-to-end ML system looks like.

𝟮/ 𝗣𝗿𝗼𝗺𝗽𝘁 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗳𝗼𝗿 𝗟𝗟𝗠𝘀 𝗯𝘆 John Berryman 𝗮𝗻𝗱 Albert Ziegler 
🔗: https://lnkd.in/dGXBYAFU

If you're obsessed with seeing the big picture and thinking in systems (like me), you'll love this book.

It will show you how to engineer your prompts effectively.

You will also learn how to manipulate your prompt, make it flexible, scalable, and squeeze the most out of your models.

P.S. As John said, this book should be called 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗳𝗼𝗿 𝗟𝗟𝗠𝘀.

𝟯/ 𝗔𝗜 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗯𝘆 Chip Huyen 
🔗: https://lnkd.in/dPKStJW3

This book has one of the best (and clearest) explanations of:

• AI vs. ML Engineering
• RAG (from classic text search to more complex, semantic-based systems)
• Building agentic systems
• LLMOps (Especially observability and user feedback)
  
𝟰/ 𝗕𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗟𝗟𝗠𝘀 𝗳𝗼𝗿 𝗣𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻 𝗯𝘆 Louis-François Bouchard 𝗮𝗻𝗱 Louie Peters 
🔗: https://lnkd.in/d9v6HyZr (𝟭𝟱% 𝗼𝗳𝗳 𝘂𝘀𝗶𝗻𝗴 𝗮𝗳𝗳𝗶𝗹𝗶𝗮𝘁𝗲 𝗰𝗼𝗱𝗲 𝗣𝗮𝘂𝗹_𝟭𝟱)

This book will walk you through how to implement all the core algorithms required to build LLM applications.

You'll use LangChain and LlamaIndex to implement all kinds of RAG techniques (starting from the vanilla architecture).

It's great for people who are mid-to-entry-level engineers.

𝟱/ 𝗟𝗟𝗠𝘀 𝗶𝗻 𝗣𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻 𝗯𝘆 Chris Brousseau 𝗮𝗻𝗱 Matthew Sharp 
🔗: https://lnkd.in/d_xqpfUz

This book focuses entirely on shipping AI.

Specifically, it does a great job of explaining:

• How to optimize LLMs using different tools
• The data engineering involved in building an LLM app
• Training and serving LLMs at scale
• How to think about infrastructure when serving LLMs
• etc.
  
𝗕𝗢𝗡𝗨𝗦: 𝗟𝗟𝗠 𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿'𝘀 𝗛𝗮𝗻𝗱𝗯𝗼𝗼𝗸 𝗯𝘆 𝗺𝗲 𝗮𝗻𝗱 Maxime Labonne 
Link: https://lnkd.in/dRmBZXdR

I couldn't list a bunch of books without mentioning my baby.

This book merges my obsession with thinking about end-to-end systems with Maxime Labonne 's love for fine-tuning.

We walk you through all the steps to build an LLM & RAG application.

Oh… and we’re working on a 2nd edition due to popular demand.

P.S. How many of these books have you read?

Senior AI Engineer • Founder @ Decoding AI ~ Building an army of AI agents and teaching you about the process.

90% of AI engineers build LLM apps the same way.

They spin up a RAG system, tweak a few prompts, and then ask:

"Does it work?"

Or worse... "Should I switch to the new model that just dropped?"

But these are the wrong questions to be asking.

Why? Because you have no way to know.

It's impossible to answer these questions without an evaluation framework.

This is why Hugo Bowne-Anderson suggests we use Evaluation-Driven Development (EDD).

The concept isn't new (especially for those with a background in ML).

The idea is to shift the development mindset from launch first, evaluate later to evaluate first, launch better.

Here's how it works:

1. 𝗨𝘀𝗲 𝘀𝘆𝗻𝘁𝗵𝗲𝘁𝗶𝗰 𝗱𝗮𝘁𝗮 to uncover failure modes before launch.
2. 𝗟𝗮𝗯𝗲𝗹 𝗯𝘆 𝗵𝗮𝗻𝗱 to build intuition about what “good” looks like.
3. 𝗕𝘂𝗶𝗹𝗱 𝗮 𝘀𝗶𝗺𝗽𝗹𝗲 𝗲𝘃𝗮𝗹𝘂𝗮𝘁𝗶𝗼𝗻 𝗵𝗮𝗿𝗻𝗲𝘀𝘀 - even a spreadsheet is enough.
4. 𝗔𝘂𝘁𝗼𝗺𝗮𝘁𝗲 it to track changes over time.
   
This loop will help you compare models, debug failures, and build confidence that your system works before it hits production.

You wouldn’t build a search engine or recommender system without robust evaluation.... 
So why build an LLM app without one?

Hugo Bowne-Anderson explains EDD beautifully in a guest post he wrote for Decoding AI Magazine.

It’s one of the most practical, no-fluff guides I’ve read on building reliable LLM systems.

Read it here → https://lnkd.in/dWpSyfcq

Digital Factory AI Lead

Reshaping industry workflows by developing AI solutions and effective agentic systems.

Lead AI Engineer

moatless

Helping industry experts and researchers to navigate the sheer endless sea of Large Language Model (LLM) research & innovation because everybody deserves to find the best information.

Stealth

I was part of a 6 month government-funded Entrepreneurship program that enabled me to work on my own projects and ideas within the LLM space.

During this time, I gained experience in building LLM-powered backends with popular in-demand tools. The prototype I ended up building included the following:

- Fetching data from an external API & processing the response (Python)
- NoSQL database (MongoDB, Postgres)
- PDF Processing (PyPDF2, PyMuPDF)
- Document Extraction (Textract/Document Transformer)
- LLM Chain with multiple steps and prompt templates (Python, Guidance/LangChain)
- API endpoint for the processed data & the LLM outputs (FastAPI)
- Cloud Infrastruture (AWS)

Since I operated on a minimal budget, efficiency and cost reduction were a priority for me. Specifically, I focused on optimizing tasks like PDF processing, document understanding, information extraction and summarization.

A1 Telekom Austria AG

During my time as a Senior at A1, I was the DS lead for the company's most advanced NLP projects.

My work included

- Reporting results to several levels of management 
- Representing cutting edge Deep Learning & NLP at the company
- Pitching these topics to the C-Level and other influential exexutives
- Giving expert talks & presentations
- Participating in workshops
- Taking part in the recruiting process for full-time Data Science internships
- Guiding selected interns and assigning them roles in my projects

For a technical overview of my work, see below.

Senior Data Scientist

I was in charge of a Social Media Listening project for Facebook, Twitter and Online Newspapers which included the following

- SQL & MongoDB databases
- Web scraper for the newspaper(s)
- Processing Facebook & Twitter API data
- Transfer Learning & Language Model Finetuning on in-domain texts
- Multiclass and Multilabel Text Classifiers based on Embeddings and Transformers (the original PyTorch implementation of BERT)
- Tableau Dashboard

On top of that, I was developing intents & entities for customer text channels, including e-mail, web search, specialized chatbots (e.g., billing), the aforementioned Social Media data and several smaller Text & Sequence Classification projects.

I then moved on to explore speech channels and became the person in charge of Speech-to-Text at the company, which included 

- Processing short IVR audio recordings
- Processing long(er) call center recordings 
- Labeling pipeline with custom Prodigy scripts
- Developing a subtitle matching algorithm for aligning audio with text & filtering out errorenous subtitles
- Developing Hybrid HMM-TDNN Speech Recognition models with Kaldi
- Switching to (back then freshly released) bleeding edge libraries & models for End-to-end Deep Learning, such as NVIDIA NeMo & Facebook FAIR
- Improving performance on Austrian dialects & in-domain recordings
- Adapting the in-house Classification models mentioned earlier for the text output of the speech channels

In addition, I was able to work on PoCs in the areas of Document Extraction, Document Classification, Entity Extraction & Question Answering.

Data Scientist

Universität Wien

Research in Cognitive & Economic Psychology:

- Dual process theory (Type I/II thinking)
- Metacognition theory
- Experimental assessment of cognitive biases
- Leveraging aggregated mouse cursor movement data

Junior Researcher

Coursera

Data Science Education

As the founder of LLM Watch, my goal is to curate the most relevant Large Language Model (LLM) research and make it accessible to a broad audience. 

I distill complex technical topics into summaries that respect your time. 

Before embarking on this journey, I led Deep Learning R&D at one of Europe's largest telecom companies. There, I spearheaded numerous AI projects with a focus on Natural Language Processing (NLP) and Automatic Speech Recognition (ASR). 

With more than 6 years of experience in AI, I was there before the hype. 

I try to offer a healthy balance between shiny new things and what really matters.

Pascal Biese

Yann LeCun's first big paper after he left Meta. Here's what they did:

For years, training Joint-Embedding Predictive Architectures (JEPAs) has been more art than science. Stop-gradients, teacher-student networks, carefully tuned EMA schedules – a delicate dance of heuristics to prevent representation collapse. But what if there was a principled way?

Self-supervised learning promised to unlock the power of unlabeled data, yet state-of-the-art methods like DINOv2 and I-JEPA rely on complex recipes with dozens of hyperparameters. Each new dataset or architecture requires extensive tuning. Even worse, the training loss often doesn't correlate with downstream performance, making model selection nearly impossible without expensive labeled evaluation.

The core problem: we've lacked a theoretical understanding of what makes good representations. LeJEPA changes this by proving a fundamental result – isotropic Gaussian embeddings uniquely minimize worst-case risk across downstream tasks. 

Tho achieve this, they introduce SIGReg (Sketched Isotropic Gaussian Regularization), a novel objective that enforces this optimal distribution using random projections and characteristic function matching. The beauty lies in its simplicity: linear complexity, bounded gradients, and provably correct by construction.

LeJEPA eliminates stop-gradients and teacher-student networks entirely while achieving competitive performance – 79% on ImageNet-1k with a frozen ViT-H/14. And here's the best part: it works out-of-the-box across 60+ architectures and diverse domains - in-domain pretraining with LeJEPA on small datasets (even just 1,000 samples) outperforms transfer learning from massive frontier models trained on billions of images.

This could fundamentally shift how we think about foundation models. Instead of relying on ever-larger generic models, practitioners in specialized domains can now train competitive models directly on their data with minimal hyperparameter tuning. 

The fact that LeJEPA's training loss actually correlates 99% with downstream performance also solves the critical model selection problem that has plagued SSL research.

What do you think? Is Yann onto something?
↓ 
𝐖𝐚𝐧𝐭 𝐭𝐨 𝐤𝐞𝐞𝐩 𝐮𝐩? Join my newsletter with 50k+ readers and be the first to learn about the latest AI research: llmwatch.com 💡

AI Lead at PwC <​/​> Daily AI highlights for 80k+ experts 📲🤗

LeJEPA: Provable and Scalable Self-Supervised Learning

Microsoft just announced the "Era of Agentic Organization". Here's what that means:

As LLMs scale their reasoning capabilities, we need new techniques: sequential thinking is too slow, and parallel thinking wastes compute by running independent traces that can't learn from each other mid-process. The industry needs a paradigm that combines the adaptivity of sequential reasoning with the speed of parallelization.

For this reason, Microsoft Research developed AsyncThink. AsyncThink is a new approach where a single LLM learns to play two roles simultaneously: an organizer that dynamically structures the reasoning process through Fork and Join operations, and workers that execute sub-queries concurrently. 

Their key contribution is that the model learns via reinforcement learning to optimize its own thinking structure - deciding when to split problems, which sub-queries to pursue, and when to merge intermediate results.

The results? 28% lower inference latency compared to parallel thinking while maintaining higher accuracy on mathematical reasoning. Models trained only on simple countdown problems successfully generalize their asynchronous thinking capabilities to completely unseen domains like Sudoku and complex math problems.

The more tasks AI is supposed to solve within an organization, the more important it becomes to structure all of this in a scalable and controllable way. AI agents need to learn to organize their reasoning like skilled project managers, adaptively coordinating multiple thinking processes based on the problem at hand. 
↓ 
𝐖𝐚𝐧𝐭 𝐭𝐨 𝐤𝐞𝐞𝐩 𝐮𝐩? Join my newsletter with 50k+ readers and be the first to learn about the latest AI research: llmwatch.com 💡

The Era of Agentic Organization

Can we train billion-parameter models without gradients? Oxford researchers just showed it's possible - at 100x the speed.

Evolution Strategies have always been an elegant alternative to backpropagation. They don't need differentiability, handle noisy objectives gracefully, and parallelize beautifully. But there's been a fundamental problem: they simply don't scale to modern model sizes.

University of Oxford's new EGGROLL algorithm mirrors what made LoRA successful for fine-tuning - instead of perturbing all parameters with full-rank matrices, use low-rank approximations. For each population member, sample matrices A and B to form a rank-r perturbation AB^T rather than a full matrix E. This cuts memory from mn to r(m+n) per layer and reduces forward pass compute proportionally.

What makes this work is subtle but powerful: while each perturbation is low-rank, the final parameter update averages across the entire population, yielding a full-rank update when the population is large enough. The researchers prove this approximation converges to full-rank ES at an O(1/r) rate - faster than typical central limit theorem rates due to the symmetry of the problem.

The results validate the approach across three domains. In pure integer pretraining of recurrent language models, EGGROLL enables stable training with population sizes exceeding 250,000 - orders of magnitude larger than prior work. For LLM reasoning tasks, it matches or exceeds GRPO while maximizing hardware utilization. And critically, it achieves near-inference throughput speeds, making gradient-free optimization genuinely practical at scale.

This opens real possibilities for training systems that can't use backpropagation: end-to-end neurosymbolic architectures, models with discrete components, or deployment scenarios where integer-only computation is essential for efficiency.
↓ 
𝐖𝐚𝐧𝐭 𝐭𝐨 𝐤𝐞𝐞𝐩 𝐮𝐩? Join my newsletter with 50k+ readers and be the first to learn about the latest AI research: llmwatch.com 💡

Evolution Strategies at the Hyperscale

We might be systematically underestimating LLMs.

We've been told that reinforcement learning is essential for frontier reasoning capabilities. Massive compute budgets, curated datasets, reward models - the whole apparatus. 

But new research from Harvard University challenges this expensive paradigm with a surprisingly simple question: can we just sample better from base models?

The paper introduces a training-free algorithm that achieves reasoning performance matching or exceeding GRPO (the standard RL method) across mathematics, coding, and science benchmarks. 

The key insight revolves around the "power distribution" - sampling from p^α instead of p. Unlike low-temperature sampling which greedily averages future token likelihoods, power sampling accounts for entire future completion paths. 

This seemingly subtle difference makes a big difference: it naturally upweights tokens with fewer but higher-likelihood futures, essentially planning ahead rather than getting trapped in dead-ends.

Using Metropolis-Hastings MCMC to approximate this distribution, the researchers show that Qwen2.5-Math-7B can reach 74.8% on MATH500 without any additional training - nearly matching the 78.5% from full RL posttraining. 

On out-of-domain tasks like HumanEval, it outperforms by up to 20 percentage points. Perhaps most remarkably, unlike RL which suffers from diversity collapse, this method maintains strong pass@k performance across multiple samples.

The takeaway is that base models may already contain the reasoning capabilities we're spending enormous resources to extract through RL. This would open up a new axis for inference-time compute scaling that's accessible to anyone with a base model - no training infrastructure, curated datasets, or verifiers required.

We're still learning how to properly query the "intelligence" that already exists in our models.
↓ 
𝐖𝐚𝐧𝐭 𝐭𝐨 𝐤𝐞𝐞𝐩 𝐮𝐩? Join my newsletter with 50k+ readers and be the first to learn about the latest AI research: llmwatch.com 💡

taplio

The Battle

Paul Iusztin

Pascal Biese

Followers

Posts

Engagements

Likes

Performance ScorePaul Iusztin's LinkedIn Performance

Performance ScorePascal Biese's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Paul Iusztin

Pascal Biese

Followers

Posts

Engagements

Likes

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Paul Iusztin's LinkedIn Performance

Performance ScorePascal Biese's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScorePaul Iusztin's LinkedIn Performance