Everseen

✅ Actively involved in all stages of ML project lifecycle.

𝐄𝐯𝐞𝐫𝐬𝐞𝐞𝐧 is an AI company that empowers end-to-end visibility across the entire supply chain for the biggest retail clients in the world, aiming to  reduce shrink and increase perpetual inventory accuracy using it’s VisualAI platform.

☑ Worked on a theft-prevention system that was showcased at 𝐍𝐞𝐰𝐘𝐨𝐫𝐤 𝐍𝐑𝐅2023, and at numerous demo-sessions that 𝐚𝐭𝐭𝐫𝐚𝐜𝐭𝐞𝐝 3 𝐧𝐞𝐰 𝐜𝐥𝐢𝐞𝐧𝐭𝐬 into wanting the product for their stores. 

☑ 𝐇𝐚𝐧𝐝𝐥𝐢𝐧𝐠 project/tooling 𝐫𝐚𝐦𝐩-𝐮𝐩 phase for new members within the team, successfully reducing the period from 1 𝐦𝐨𝐧𝐭𝐡 𝐭𝐨 𝐥𝐞𝐬𝐬 𝐭𝐡𝐚𝐧 1 𝐰𝐞𝐞𝐤.

☑ Handled 𝐞𝐧𝐝-𝐭𝐨-𝐞𝐧𝐝 engineering and 𝐝𝐞𝐩𝐥𝐨𝐲𝐦𝐞𝐧𝐭 to 5 𝐩𝐢𝐥𝐨𝐭 𝐬𝐭𝐨𝐫𝐞𝐬.

☑ Improved the 𝐢𝐭𝐞𝐫𝐚𝐭𝐢𝐨𝐧 𝐬𝐩𝐞𝐞𝐝 within the team by 𝐚𝐮𝐭𝐨𝐦𝐚𝐭𝐢𝐧𝐠 𝐌𝐋𝐎𝐩𝐬 practices such as Structured Model Experimenting, Resources Monitoring, Model Registry, CI/CD and Jira Integrations.

☑ 𝐎𝐫𝐠𝐚𝐧𝐢𝐳𝐞𝐝 and established a 𝐝𝐨𝐜𝐮𝐦𝐞𝐧𝐭𝐚𝐭𝐢𝐨𝐧 𝐬𝐭𝐫𝐮𝐜𝐭𝐮𝐫𝐞 which enabled us to 𝐫𝐞𝐚𝐜𝐡 > 70% 𝐜𝐨𝐯𝐞𝐫𝐚𝐠𝐞 both on code,tools documentation and custom models tailored for various clients.

☑ Worked on a live camera feed processing pipeline that ran 3 𝐝𝐞𝐞𝐩 𝐥𝐞𝐚𝐫𝐧𝐢𝐧𝐠 𝐦𝐨𝐝𝐮𝐥𝐞𝐬 while achieving 24-25 FPS on a 4K video stream with high QPS rate.

☑ Implemented a 𝐦𝐨𝐧𝐢𝐭𝐨𝐫𝐢𝐧𝐠 𝐩𝐢𝐩𝐞𝐥𝐢𝐧𝐞 𝐟𝐨𝐫 𝐨𝐮𝐫 𝐞𝐝𝐠𝐞 𝐝𝐞𝐩𝐥𝐨𝐲𝐦𝐞𝐧𝐭𝐬 that saved time with hardware requirements and optimisations using Prometheus, Grafana, Kibana and ES.

☑ Actively 𝐩𝐚𝐫𝐭𝐢𝐜𝐢𝐩𝐚𝐭𝐢𝐧𝐠 𝐢𝐧 𝐭𝐡𝐞 𝐥𝐢𝐟𝐞𝐜𝐲𝐜𝐥𝐞 of the current project, providing support to QA, Development and Deployment processes.

☑ 𝐀𝐮𝐭𝐨𝐦𝐚𝐭𝐞𝐝 the multi-gpu 𝐦𝐨𝐝𝐞𝐥 𝐜𝐨𝐧𝐯𝐞𝐫𝐬𝐢𝐨𝐧 and integration with high-throughput Model Serving technologies like Nvidia Triton Server.

Senior Machine Learning Engineer

Neural Bits

✅ Writing the Neural Bits Newsletter to help AI/ML Engineers upskill in their career.
☑ Making AI accessible to a larger general audience by explaining concepts in simple terms.
☑ Sharing insights and practical advice on AI/ML Systems.

Author

Decoding ML

Sharing insights and actionable advice towards building production-ready ML systems.

Co-Founder

HandsOnLLMs

✅ Contributed to a free hands-on course on end-to-end LLMops.

Worked alongside 𝐏𝐚𝐮 𝐋𝐚𝐛𝐚𝐫𝐭𝐚 𝐁𝐚𝐣𝐨 𝐚𝐧𝐝 𝐏𝐚𝐮𝐥 𝐈𝐮𝐬𝐳𝐭𝐢𝐧 into developing this end-to-end course to help students 𝐛𝐮𝐢𝐥𝐝 𝐚 𝐜𝐨𝐦𝐩𝐥𝐞𝐭𝐞 𝐋𝐋𝐌 𝐚𝐩𝐩𝐥𝐢𝐜𝐚𝐭𝐢𝐨𝐧 while following the 𝐛𝐞𝐬𝐭 𝐩𝐫𝐚𝐜𝐭𝐢𝐜𝐞𝐬 both in software development, model deployment and MLOps integrations.

☑ Implemented the EDA stage on LLM fine-tuning dataset.
☑ Implemented the LangChain logic and modules that connected the QDrant vector database to inference pipeline and model registry.
☑ Worked on the training/fine-tuning pipeline and prompt monitoring using CometML.
☑ Provided support and approved Pull Requests.
☑ Implemented and documented best practices on Prompt Templates and LLM workflow.

Check the course here: http://bit.ly/3Rxjyvp

Co Author

IVFuture

✅ Worked on on a suite of tools including Generative AI, LLM’s and Vision Models.

𝐈𝐕𝐅𝐮𝐭𝐮𝐫𝐞’𝐬 mission is to transform common businesses, communities and people’s needs into intelligent and innovative software solutions, to simplify life through scalable technology.

☑ Implemented a 𝐝𝐞𝐦𝐨 for virtual try-on fashion tool called 𝐓𝐀𝐈𝐒, which was presented to investors and attracted interest from various fashion providers across Romania.

☑ Implemented a CV solution, that was able to accurately showcase over 25 𝐛𝐨𝐝𝐲 𝐦𝐞𝐚𝐬𝐮𝐫𝐞𝐦𝐞𝐧𝐭𝐬 𝐢𝐧 3𝐃 using 2D cameras, with an 𝐞𝐫𝐫𝐨𝐫-𝐫𝐚𝐭𝐞 𝐨𝐟 5-10% while using cheap hardware, <700$ 𝐩𝐞𝐫 𝐬𝐞𝐭𝐮𝐩, considering industry available setups 𝐞𝐱𝐜𝐞𝐞𝐝 10.000$.

☑ Developed and integrated an 𝐋𝐋𝐌 13𝐁 𝐜𝐡𝐚𝐭 𝐚𝐠𝐞𝐧𝐭, fine-tuned on specific descriptions that was used by the team to automatically 𝐥𝐚𝐛𝐞𝐥 𝐡𝐮𝐧𝐝𝐫𝐞𝐝𝐬 𝐨𝐟 𝐩𝐫𝐨𝐝𝐮𝐜𝐭𝐬.

☑ Fine-tuned 𝐒𝐭𝐚𝐛𝐥𝐞 𝐃𝐢𝐟𝐟𝐮𝐬𝐢𝐨𝐧1.5 and tailored it to specific product generation use-case.

☑ Worked on a solution for 𝐓𝐓𝐒 𝐚𝐧𝐝 𝐒𝐓𝐓 and exposed it as an API, which was able to accurately 𝐜𝐥𝐨𝐧𝐞 𝐯𝐨𝐢𝐜𝐞𝐬 𝐟𝐫𝐨𝐦 𝐚 𝐬𝐡𝐨𝐫𝐭 5𝐬𝐞𝐜 𝐚𝐮𝐝𝐢𝐨 𝐜𝐥𝐢𝐩 and synthesize text-to-speech and speech-to-text.

☑ Worked on a 𝐑𝐀𝐆 𝐚𝐠𝐞𝐧𝐭, that was able to ingest document context and provide assistant-like answers on user queries.

Machine Learning Consultant - Generative AI

Odeen Software

✅ Worked on Flowstate, a Computer Vision solution developed for 𝐔𝐑𝐁𝐍𝐒𝐔𝐑𝐅 𝐌𝐞𝐥𝐛𝐨𝐮𝐫𝐧𝐞, 𝐀𝐮𝐬𝐭𝐫𝐚𝐥𝐢𝐚, that generates 𝐬𝐨𝐜𝐢𝐚𝐥-𝐦𝐞𝐝𝐢𝐚 𝐫𝐞𝐞𝐥𝐬 𝐚𝐧𝐝 𝐬𝐡𝐨𝐫𝐭𝐬 for park attendees.

𝐎𝐝𝐞𝐞𝐧 is a company that provides IT services, staff augmentation and SASS solutions on various technology stacks, including Machine Learning, Web Development and more.

☑ Implemented a 𝐦𝐮𝐥𝐭𝐢-𝐜𝐚𝐦𝐞𝐫𝐚 𝐭𝐫𝐚𝐜𝐤𝐢𝐧𝐠 feature using SOTA Centroid ReID to track and re-identify person detections in multiple camera views.

☑ Integrated 𝐅𝐀𝐈𝐒𝐒 𝐕𝐞𝐜𝐭𝐨𝐫 𝐃𝐚𝐭𝐚𝐛𝐚𝐬𝐞 to store track embeddings and perform ReID searching for similar tracks in a batch fashion.

☑ Used 𝐃𝐞𝐞𝐩𝐅𝐚𝐜𝐞 to detect and group track detections with face detections, and human pose prediction to filter accurate poses to consider when generating reels.

☑ Worked on the video streaming workflow, offline batch tracking and edge integration with Viso.AI.

Machine Learning Consultant

Upwork

💎 Top Rated • 💯 Job Success Rate • ⭐⭐⭐⭐⭐ Reviews

✅ Developed plug-and-play ML/CV modules.

☑ Client had over 50𝐤 𝐫𝐞𝐜𝐞𝐢𝐩𝐭𝐬 from various stores, and needed to automate the process of analyzing these receipts. I’ve implemented a 𝐌𝐋 𝐬𝐨𝐥𝐮𝐭𝐢𝐨𝐧 𝐬𝐚𝐯𝐢𝐧𝐠 𝐮𝐩 𝐭𝐨 80% of the time used to detect, decode and organise these receipts by their QR code metadata.

☑ Client wanted a 𝐄𝐝𝐮𝐓𝐞𝐜𝐡 𝐰𝐞𝐛 𝐩𝐥𝐚𝐭𝐟𝐨𝐫𝐦 that allows parents to create quizz tests for their children based on an uploaded document (.pdf, .docx). I’ve implemented a full-stack web application with included ML components which loaded a document, applied ETL, generated quizz test and construct the UI view.

☑ Client wanted a CV plugin that would detect various icons from a phone screen, classify them and automatically create app testing workflows. Using template matching, pyramid-detection and super-resolution GAN’s, I’ve managed to 𝐨𝐩𝐭𝐢𝐦𝐢𝐬𝐞 𝐚𝐧𝐝 𝐚𝐮𝐭𝐨𝐦𝐚𝐭𝐞 𝐜𝐥𝐨𝐬𝐞 𝐭𝐨 100% on a few use-cases, and around 30-40% 𝐨𝐧 𝐭𝐡𝐞 𝐫𝐞𝐬𝐭 of the use-cases required.

☑ 𝐂𝐨𝐧𝐬𝐮𝐥𝐭𝐞𝐝 on various Computer Vision topics and project ideas.

Top Rated Machine Learning Engineer

✅ Developed and deployed CV/ML modules into existing API's.

☑ Client wanted a CV solution that 𝐩𝐫𝐨𝐜𝐞𝐬𝐬𝐞𝐬 𝐬𝐩𝐨𝐫𝐭𝐬 𝐠𝐚𝐦𝐞𝐬 𝐚𝐧𝐝 𝐬𝐞𝐧𝐝𝐬 𝐧𝐨𝐭𝐢𝐟𝐢𝐜𝐚𝐭𝐢𝐨𝐧𝐬 𝐰𝐡𝐞𝐧𝐞𝐯𝐞𝐫 𝐭𝐡𝐞 𝐬𝐜𝐨𝐫𝐞 𝐜𝐡𝐚𝐧𝐠𝐞𝐬. whenever the score changes. I’ve applied smart-template matching, targeted motion detection and OCR to extract the team names, scores and send a push message whenever the score changed.

☑ Client wanted a CV solution that scans documents for hand signatures, identifies faulty ones and groups them in a specific structure. Solution was based on deep learning and OCR post-processing, deployed as a docker ML service which helped the client to 𝐬𝐩𝐞𝐞𝐝-𝐮𝐩 𝐭𝐡𝐞 𝐩𝐫𝐨𝐜𝐞𝐬𝐬 𝐛𝐲 𝐚 3𝐱 𝐟𝐚𝐜𝐭𝐨𝐫, relying far less on Humans-in-the-Loop activity.

☑ Client wanted a project concept report on how to apply custom texture from an image onto a 3D model. I’ve researched and planned a workflow that included Object Detection, Instance Segmentation and UV mapping.

Machine Learning Engineer

Elektrobit

✅ Developed multiple Computer Vision solutions and deployed them on edge devices.

𝐄𝐥𝐞𝐤𝐭𝐫𝐨𝐛𝐢𝐭 is key supplier of embedded software solutions for the automotive industry.

☑ Worked on a 𝐣𝐨𝐢𝐧𝐭-𝐩𝐫𝐨𝐣𝐞𝐜𝐭 𝐰𝐢𝐭𝐡 𝐁𝐌𝐖 𝐌𝐮𝐧𝐢𝐜𝐡 𝐟𝐨𝐫 𝐄𝐏𝐈 (European Processor Initiative 2020) and deployed 𝐚 𝐜𝐚𝐫 𝐝𝐞𝐭𝐞𝐜𝐭𝐢𝐨𝐧 𝐚𝐧𝐝 𝐩𝐚𝐭𝐡-𝐩𝐥𝐚𝐧𝐧𝐢𝐧𝐠 𝐬𝐨𝐥𝐮𝐭𝐢𝐨𝐧 on an Nvidia Jetson AGX board, placed in a real 𝐁𝐌𝐖 𝐱5 𝐭𝐞𝐬𝐭 𝐜𝐚𝐫, which was able to detect and map the environment around the test vehicle using video feed from 2xIntelCameras and 1xLidarSensor.

☑ 𝐒𝐮𝐜𝐜𝐞𝐬𝐬𝐟𝐮𝐥𝐥𝐲 𝐩𝐮𝐛𝐥𝐢𝐬𝐡𝐞𝐝 𝐚 𝐬𝐜𝐢𝐞𝐧𝐭𝐢𝐟𝐢𝐜 𝐩𝐚𝐩𝐞𝐫 titled GFPNet, under the supervision of the Head of AI @Elektrobit Romania, at the 𝐈𝐄𝐄𝐄 𝐑𝐨𝐛𝐨𝐭𝐢𝐜𝐬 𝐚𝐧𝐝 𝐀𝐮𝐭𝐨𝐦𝐚𝐭𝐢𝐨𝐧 𝐋𝐞𝐭𝐭𝐞𝐫𝐬 2020, within the Machine Learning and Robotics field that studied the shape reconstruction of occluded 3D objects.

☑ Generated synthetic data using 𝐂𝐚𝐫𝐥𝐚𝐒𝐢𝐦𝐮𝐥𝐚𝐭𝐨𝐫 (𝐔𝐄4) to train deep neural networks for specific autonomous driving scenarios, improving the accuracy of certain prediction.

☑ Worked on development and optimisation of a vehicle direction and position classifier neural network, integrated directly on the Renesas 𝐑-𝐂𝐚𝐫-𝐇3𝐄 𝐜𝐡𝐢𝐩 (𝐒𝐨𝐂), further deployed to one of the clients test car.

☑ Worked on a 𝐦𝐨𝐝𝐞𝐥 𝐜𝐨𝐧𝐯𝐞𝐫𝐬𝐢𝐨𝐧 𝐚𝐧𝐝 𝐝𝐞𝐩𝐥𝐨𝐲𝐦𝐞𝐧𝐭 𝐩𝐢𝐩𝐞𝐥𝐢𝐧𝐞 for edge devices, that allowed a model to be ported in Python/C++ with no further intervention, 𝐬𝐩𝐞𝐞𝐝𝐢𝐧𝐠 𝐮𝐩 𝐭𝐡𝐞 𝐩𝐫𝐨𝐜𝐞𝐬𝐬 𝐛𝐲 𝐨𝐯𝐞𝐫 50%.

✅ Researched computer vision topics, mainly targeted towards autonomous driving scenarios.

☑ Researched 𝐇𝐲𝐝𝐫𝐚𝐍𝐞𝐭𝐬 for 𝐚𝐮𝐭𝐨𝐦𝐨𝐭𝐢𝐯𝐞 𝐯𝐢𝐬𝐢𝐨𝐧 𝐭𝐚𝐬𝐤𝐬, using ResNet (50, 100) encoder backbones and multiple prediction heads (classification, 3D bounding-box, object detection and orientation), the research 𝐫𝐞𝐩𝐨𝐫𝐭 𝐛𝐞𝐢𝐧𝐠 𝐟𝐮𝐫𝐭𝐡𝐞𝐫 𝐮𝐬𝐞𝐝 to develop, train and deploy a model on a small 𝐟𝐨𝐮𝐫-𝐰𝐡𝐞𝐞𝐥𝐞𝐝 𝐫𝐨𝐛𝐨𝐭.

☑ Researched 𝐏𝐨𝐢𝐧𝐭𝐂𝐥𝐨𝐮𝐝 𝐬𝐞𝐠𝐦𝐞𝐧𝐭𝐚𝐭𝐢𝐨𝐧 𝐟𝐫𝐨𝐦 𝐃𝐞𝐩𝐭𝐡 using a X-Box KinectSensor, feature that was further tested in a stack of models 𝐨𝐩𝐞𝐫𝐚𝐭𝐢𝐧𝐠 𝐨𝐧 𝐚 𝐟𝐥𝐲𝐢𝐧𝐠 𝐝𝐫𝐨𝐧𝐞 that followed QR markers on the floor.

☑ Worked on a model-development dashboard that 𝐚𝐥𝐥𝐨𝐰𝐞𝐝 𝐮𝐬𝐞𝐫𝐬 𝐭𝐨 𝐮𝐩𝐥𝐨𝐚𝐝 𝐝𝐚𝐭𝐚, 𝐛𝐮𝐢𝐥𝐝 𝐭𝐡𝐞 𝐦𝐨𝐝𝐞𝐥, 𝐬𝐜𝐡𝐞𝐝𝐮𝐥𝐞 𝐚 𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐣𝐨𝐛, inspect metrics and download artifacts on specific targets (.pt, .onnx, .trt, .gzip).

Junior Machine Learning Engineer

Universitatea Transilvania din Brașov

✅ Teaching 𝐢𝐧𝐭𝐞𝐫𝐦𝐞𝐝𝐢𝐚𝐭𝐞 programming concepts and intro to 𝐦𝐚𝐜𝐡𝐢𝐧𝐞 𝐥𝐞𝐚𝐫𝐧𝐢𝐧𝐠.

☑ Java OOP, Python, Git
☑ Data Structures and Algorithms, Graphs
☑ Artificial Neural Networks
☑ Supervised and unsupervised learning
☑ House Price Prediction
☑ Activation Functions
☑ Hand Digits Recognition using Computer Vision and Deep Learning

Teaching Assistant

Siemens

✅ Platform to manage/schedule an industrial process workflow.

𝐒𝐢𝐞𝐦𝐞𝐧𝐬 𝐂𝐨𝐫𝐩𝐨𝐫𝐚𝐭𝐞 𝐓𝐞𝐜𝐡𝐧𝐨𝐥𝐨𝐠𝐲 is a strategic partner to Siemens operative units. It provides important services along the entire value chain – from research and development to production and quality assurance, as well as optimized business processes.

☑ 𝐏𝐥𝐚𝐧𝐧𝐞𝐝 𝐭𝐡𝐞 𝐜𝐨𝐦𝐩𝐨𝐧𝐞𝐧𝐭𝐬 for an app that would allow users to inspect and manage stock for a potato-chips factory.

☑ Implemented the 𝐬𝐞𝐜𝐮𝐫𝐢𝐭𝐲 𝐥𝐚𝐲𝐞𝐫 for admin, observer, user with different functionalities for each category.

☑ 𝐈𝐦𝐩𝐥𝐞𝐦𝐞𝐧𝐭𝐞𝐝 𝐭𝐡𝐞 𝐟𝐮𝐥𝐥-𝐬𝐭𝐚𝐜𝐤 𝐥𝐨𝐠𝐢𝐜, communications between components and frontend dashboard.

☑ 𝐖𝐨𝐧 3’rd place amongst over 10 projects presented.

Student Internship

Hard work always pays off, be consistent!

I'm a Senior Machine Learning Engineer with a keen interest in developing AI/ML solutions for real world problems and helping others get started on their Machine Learning Journey.

Let's connect @:
📘 https://medium.com/@alexandrurazvant
✉️ alexandrurazvant@gmail.com
🌐 https://www.neuraleaps.com

Alex Razvant

Want to learn about Agents, MCP, Multimodal AI, and LLMOps?

Miguel Otero and I have released an open-source course where we build an Multimodal Agent that can understand videos, step by step.

The project idea is simple: 𝐘𝐨𝐮 𝐮𝐩𝐥𝐨𝐚𝐝 𝐚 𝐯𝐢𝐝𝐞𝐨, 𝐚𝐧𝐝 𝐭𝐡𝐞𝐧 𝐜𝐡𝐚𝐭 𝐰𝐢𝐭𝐡 𝐢𝐭..

However, there are many complex challenges, underlying this simple idea.

Some of them are:
- How to build and debug MCP Servers in Python
- Understanding Multimodal Ingestion (video, audio, text, images)
- Designing efficient FastAPI APIs, Background Tasks
- Tool Calling / Router Agents from scratch, in Python
- Understanding Vision Language Models (VQA), Audio Transcription Models (STT)
- How to connect APIs (FastAPI) with UIs (React JS)

Throughout the course,  we explain each module with an general audience pace in mind.

We aimed for it to be easy to follow, easy to understand.

However, we still kept it technical enough.

You don't have to register or be an AI Expert to take this course. It's free.

The course lessons split:
1/ Course Overview
🔗 https://lnkd.in/dm73KT_i

2/ Understand Multimodal Pipelines
🔗  https://lnkd.in/dmn_RBfm

3/ Advanced FastAPI and Pydantic (extra)
🔗  https://lnkd.in/drpKtPFD

4/ Understand the MCP Protocol (extra)
🔗  https://lnkd.in/dJzNp8mH

5/ Build an MCP Server for Multimodal Ingestion
🔗  https://lnkd.in/d33Rsm4e

6/ Building a Tool Calling Agent from Scratch
🔗  https://lnkd.in/dDz6-zZU

7/ Building the full-stack (API, Agent, UI) Application
🔗 https://lnkd.in/dJezq4ry 

8/ Kubrick: The Multimodal Agent Round-up
🔗 https://lnkd.in/drZWiBdT 

Find the GitHub Repository here (give us a ⭐ ):
🔗 https://lnkd.in/dudc7BZS


Tech Stack:
- FastAPI, React JS, FastMCP, MCP Inspector, Docker Compose
- LLMOps with Opik (by @Comet)
- Multimodal Processing with @Pixeltable
- LLMs, VLMs and STT (Speech To Text) AI Models.

What you'll learn as a result:
- Build an MCP Server
- Build an API in Python with FastAPI
- Add LLM/Agent observability
- Process video/audio multimodal data.
- Connect multiple components into an AI System

Have fun learning & building!
✌️

Senior AI Engineer | Author @ NeuralBits | Helping engineers understand, build, and master AI.

CNNs remember where. ViTs remember what  

In a majority of Computer Vision models, the feature extractors (i.e, backbone networks) were built either as CNN encoders, ResNets, or Pyramid Networks.

Regardless of how large the Backbone is, CNNs are not that good at learning full-image context, due to how Receptive Fields work.

```
In deep learning, a receptive field is the specific area of the input data that influences the calculation of a single output feature or neuron in a CNN layer. 
```

1/ CNNs have inductive biases baked into their design
2/ Local receptive fields, hard to scale to High-Res images.
3/ Hierarchical feature extraction
4/ This can be improved, but at the cost of the model's depth

In the top section of the attached image, the `blue-square trace represents a receptive field from High-Level features onto Low-Level features`.

The red ❌  signs represent the feature regions that are not covered by the Receptive Field.

Thus, with CNNs, a larger image context is left out.

---

ViTs, on the other hand, see an image as a sequence of patches and mix them globally through self-attention. 

When feeding an image to a transformer, we first split it into patches of NxN pixels.
Each patch receives a positional encoding vector.
The resulting embedding vector is passed through the Transformer layers.

That makes Transformers:
1/ Permutation-invariant, positional embeddings are required.
2/ Focus on learning global + local level features 
3/ Adversarial Robustness (due to global-context, can spot tinkered features)
3/ Attention Mechanism over patches is computed in Many-to-Many order.
4/ Better for tasks that require image-wide context.

To summarize:
CNNs are limited in capturing global context due to their Receptive Fields.
ViTs compute relationships with and between each patch in the image.

CNNs are better for locality-focused tasks, e.g, Object Detection.
ViTs are better for context-focused tasks, e.,g VQA.

♻️ Reshare and let others learn this too.
Follow me to learn more about AI Engineering and production-ready AI Systems!

"What should I know about NVIDIA AI Stack?"

Here's my answer, tailored for those starting out in AI regarding NVIDIA's ecosystem.

𝗧𝗵𝗲 𝗘𝗰𝗼𝘀𝘆𝘀𝘁𝗲𝗺
NVIDIA doesn't just build the GPUs that power a large chunk of AI Infrastructure.

They've built a full end-to-end ecosystem.
> 𝙃𝙖𝙧𝙙𝙬𝙖𝙧𝙚: All the architectures and iterations (Turing, Ada, Hopper, Blackwell, etc.)
> 𝙎𝙤𝙛𝙩𝙬𝙖𝙧𝙚: CUDA, cuDNN, TensorRT, Triton, NeMo, Dynamo, and more.

This stack powers everything from training LLMs to deploying models at the edge, across various fields (Robotics, Vision, Language, Speech, Agentic and Embodied AI).

But for a beginner, these are too many, each with its place in the end-to-end AI workflow.
It's easy to get lost.

So, here's my advice, split into 3 categories:
1/ General Important Details
2/ Engineering Advice
3/ Software Stack

𝗚𝗲𝗻𝗲𝗿𝗮𝗹 𝗜𝗺𝗽𝗼𝗿𝘁𝗮𝗻𝘁 𝗗𝗲𝘁𝗮𝗶𝗹𝘀
First, I recommend that you learn at a very high level the GPU components.

This will get you started in understanding the hardware black box and differentiating between various GPU/DSA architectures.

Learn about:
1. GPU Cores
What are they? What's the difference between a CUDA Core and a Tensor Core?

2. Memory Hierarchy
Differentiate between VRAM, SRAM, and Registers. This will open doors to understanding FlashAttention, which is a key optimization in LLMs.

3. Precision & Data Types
The majority of LLMs are trained in FP32 or Mixed (FP32/FP16). For Inference, FP16/BF16,  FP4. This will also help you understand the quantization data types.

𝗘𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝗔𝗱𝘃𝗶𝗰𝗲
These will help you understand how to work more efficiently with NVIDIA's hardware and software stack.

1. Learn to read a GPU Cheatsheet (details about a GPU's performance)
2. Learn to use the Framework Compatibility Matrix (mandatory for SWEs, AIEs)
3. Learn to use NVIDIA's pre-built containers (built in robust, isolated environments)

𝗧𝗵𝗲 𝗦𝗼𝗳𝘁𝘄𝗮𝗿𝗲 𝗦𝘁𝗮𝗰𝗸
We have a lot on this layer, but for starters, just learn what each of these does:

1. 𝘊𝘰𝘳𝘦 > CUDA & cuDNN, the backbone for deep learning, both training and inference.
2. 𝘋𝘢𝘵𝘢 𝘗𝘳𝘰𝘤𝘦𝘴𝘴𝘪𝘯𝘨 > DeepStream for real-time video pipelines, cuDF for Dataframes on GPU.
3. 𝘐𝘯𝘧𝘦𝘳𝘦𝘯𝘤𝘦 > TensorRT/LLM for optimizing models, Triton, and Dynamo for deployment.
4. 𝘎𝘦𝘯𝘈𝘐 > Nemotron Models, NeMo Framework to create, customize, and deploy GenAI models.
5. 𝘌𝘯𝘵𝘦𝘳𝘱𝘳𝘪𝘴𝘦 > NIM for batteries-included GenAI deployments. Riva for Speech.
6. 𝘌𝘥𝘨𝘦 𝘈𝘐 > Jetson Xavier, DGX Spark, Orin platforms.
7. 𝘚𝘤𝘢𝘭𝘪𝘯𝘨 > K8S GPU Operator and KAI Scheduler for GPU resource management at scale.

---
There are a few more, but this is a solid start.

In the end, if you spend a few hours reading through these, you'll get a good overview of what's out there.

♻️ Share this post, and research it later.

No BS AI/ML Content | ML Engineer with a Plot Twist 🥷30M+ Views 📝

Done-for you Audience Building & Content Writing for AI/ML Founders that frees up 100% of your time, so you can focus entirely on building your company  
  
Is your work too important to ignore? 💌 https://tally.so/r/wk69VJ  (or shoot me a DM) 👉
  
It's 2025 and we're at the peak of the attention economy  
  
Where Authenticity is king, visibility is currency and people wanna buy from people

Not faceless dehumanized brands  
  
So, they told you LinkedIn was "free marketing" you slapped some posts and called it strategy  
  
You've poured hours crafting the 'perfect' post, only to watch it sink without a trace..  
  
Then it dawned on you, your technical brilliance isn't enough to draw a crowd  
  
Hell, maybe you've even tried mimicking influencers, scrambling for any traction..  
  
Your post got 500 f*cking views... from your friends...  
  
Being ignored is soul-crushing  
  
It's not your fault tough  
  
When you started your company, you didn't sign-up for a degree in an Audience Building, Copywriting and LinkedIn strategy 

And here you are, sacrificing yet another late night for your vision  

"Am I just sabotaging my efforts?"  
  
"Is my life-work meaningless?"  
  
Getting traction feels impossible  
  
You hoped hiring some marketer would save your weekends but..  
  
Tech is a code they can't crack  
  
They can only churn out the lame, soul-sucking crap we all ignore on sight  
  
With armies of founders flooding feeds, your voice dissolves into meaningless white noise  
  
But in the process, you've wasted more of your precious time explaining 101 tech concept and rewriting low-quality drafts 

Done being ignored?  
  
Let's make the world listen 
  
I can help you with:  

✒️ No BS AI/ML Content  
⭐ LinkedIn Profile Revamp
🎯 Website Copy Optimization  

I've done it before:

↳ For founders backed by YC, a16z, Sequoia, and Techstars
↳ For bootstrappers in LLMs, ML, Vector DBs, Data Analytics, and Engineering
↳ For AI and Machine Learning leaders at Google, Microsoft, Airbnb, and YouTube

If they could rise above the noise, so can you

Yeah, even if you're an introvert with a "boring" niche

Still with me?  
  
Amazing—You're amazing!  
  
Cheers,  
Paolo

My booking link: 💌 https://tally.so/r/wLW0oz (or shoot me a DM) 👉

Paolo Perrone

NVIDIA just made their best AI models free.

99% of developers have no idea.

While you burn cash on OpenAI credits...
NVIDIA open-sourced Nemotron.
Their entire agentic AI family 🤯

Free Models That Actually Ship:
• Nano: Runs on your laptop
• Super: Single GPU beast
• Ultra: Multi-GPU monster

What They Do:
→ Graduate-level reasoning
→ Math without hallucinations
→ Code that compiles
→ Visual reasoning that works
→ Tool calling. No loops.

The genius part:

They published EVERYTHING.
Models. Datasets. Techniques.
Actually open source.

Game Changers:
• 6x faster on new hardware
• No API limits
• Runs on a Raspberry Pi
• Commercialize it. No strings.

The uncomfortable truth:

You're waiting for GPT-5.
Smart teams ship with Nemotron.

Who's using it:
- Accenture
- CrowdStrike  
- ServiceNow
- Deloitte

The kicker:

Download from Hugging Face. Now.
Deploy tomorrow.
Zero license fees.

Forever.

They built enterprise NIM too.
(That costs money. Models don't.)

Still paying for API credits? 

♻️ Repost before OpenAI charges per thought

No BS AI/ML Content | ML Engineer with a Plot Twist 🥷50M+ Views 📝

I just finished Google's biggest AI Agents course ever.

250+ pages of production insights. 100% free.

Here's everything they released:

DAY 1️⃣ INTRODUCTION TO AI AGENTS (50 pages)

Most agent demos fail in production.
This covers the 5 levels of AI agents, from basic reasoning to self-evolving systems, plus security and architecture patterns that scale.
🔗 https://lnkd.in/gN9JaXgR

DAY 2️⃣ AGENT TOOLS AND MCP (50 pages)

MCP servers can add new tools without your approval.
Your poetry bot suddenly gains purchasing power overnight.
This breaks down tool design, the Model Context Protocol, and hidden security risks.
🔗 https://lnkd.in/gRWDSjBr

DAY 3️⃣ CONTEXT ENGINEERING AND MEMORY (70 pages)

Agents with zero long-term memory treat every session like meeting you for the first time.
Learn how to build memory systems that actively curate context, not just save conversations.
🔗 https://lnkd.in/gbgWJ4kH

DAY 4️⃣ AGENT QUALITY (50 pages)

Traditional QA doesn't work for agents. You can't unit test judgment.
This covers the 4 pillars of agent quality (effectiveness, efficiency, robustness, safety) and how to evaluate reasoning.
🔗 https://lnkd.in/g6DHcZDx

DAY 5️⃣ PROTOTYPE TO PRODUCTION (40 pages)

Building an agent takes 2 weeks.
Trusting it in production takes 6 months.
Learn evaluation gates, circuit breakers, and evolution loops that close the gap.
🔗 https://lnkd.in/gYTakhAK

Huge props to Antonio Gulli and Shubham Saboo for making this happen, and make it 100% free

250+ pages. Zero cost. All production-focused.

♻️ Repost to save a fellow AI engineer 6 months of trial and error

4 math concepts you need to understand LLMs in 10 minutes.

No PhD required.

I spent 6 months confused by AI papers.
Then I found this breakdown by Giles Thomas.
Everything clicked.

Here's the cheat sheet:

1️⃣ Vectors = Direction + Distance in Space

Think GPS coordinates, but in 50,000 dimensions.
Each position represents a different meaning.
"Cat" lives near "dog" in this space.
"Airplane" lives far away.

That's it. That's embeddings.

2️⃣ Matrices = Space Transformers

Remember rotating shapes in geometry class?
Same thing, but with meaning.

A matrix takes your 50,257-dimensional vocab
Projects it down to 768 dimensions
Then back up again.

It's compression without losing the important stuff.

3️⃣ Dot Product = Similarity Score

Multiply matching positions. Add them up.
Big number = similar vectors.
Small number = different vectors.

This is how LLMs know "king - man + woman = queen"

4️⃣ Neural Network Layers = Just Matrix Math

Remove the activation function and what's left?
Matrix multiplication.
A projection from one space to another.

Add activation back = now it can learn patterns.

The GPT-2 Example That Makes It Real:

Input: "The cat sat on the"
1. Convert each word to a one-hot vector (50,257 dimensions)
2. Project down to embeddings (768 dimensions)  
3. Let attention mechanism figure out relationships
4. Project back up to vocab space
5. Pick highest probability: "mat"

That's literally how ChatGPT works.
Just bigger.

Why This Matters:

Everyone's scared of AI math.
But it's high school algebra in a trench coat.

Once you see through the intimidation,
you can actually build stuff.

♻️ Repost to demystify AI for someone else

Sebastian Raschka, PhD just exposed the biggest secret in AI.

After dissecting 9 flagship LLMs, he found something nobody wants to admit:

We haven't invented a new architecture in 7 years.

The evidence is damning:
• DeepSeek V3: 671B parameters, still a transformer
• Kimi K2: 1 TRILLION parameters, still a transformer
• GPT-OSS: OpenAI's latest, still a transformer
• Llama 4: Meta's flagship, still a transformer

Every "breakthrough" model?
Same skeleton. Different makeup.

Here's what's actually happening:

DeepSeek V3/R1: Compressed attention (MLA) + expert routing (MoE)
Gemma 3: Sliding windows to fake longer context
SmolLM3: Removed position embeddings entirely (!)
Kimi K2: Just made DeepSeek bigger

The brutal reality:

• No new architectures since "Attention is All You Need" (2017)
• Every innovation = efficiency hack
• We're polishing, not inventing

671 billion parameters.
1 trillion parameters.
Same. Damn. Architecture.

While everyone's chasing parameter counts, Sebastian mapped the actual mutations:
• MoE: Use 37B of your 671B parameters
• Sliding attention: Pretend you have infinite context
• MLA: Compress your KV cache and pray

This isn't innovation.
It's optimization theater.

The most honest technical analysis you'll read this year.

Read it. Then ask yourself:
If we can't innovate past transformers, what happens when we hit the parameter wall?

♻️ Repost if you appreciate actual technical depth over hype

The best AI engineers master PyTorch. The rest just import it.

Here's how to join the first group:

1️⃣ Master the Foundation
🔸 NumPy arrays → PyTorch tensors: https://lnkd.in/grmapDsK
🔸 Tensor operations that matter: https://lnkd.in/gjk3ErXu

2️⃣ Learn PyTorch Basics (Skip the Fluff)
🔸 60-minute blitz (actually 60 mins): https://lnkd.in/gU86ZGqt
🔸 What torch.nn really does: https://lnkd.in/gEmCGg6R
🔸 Daniel Bourke's 1-day crash course: https://lnkd.in/gCdBG_jc

3️⃣ Understand Autograd (The Magic Behind It All)
🔸 Karpathy's micrograd repo: https://lnkd.in/g_hTUmji
🔸 Karpathy's micrograd video: https://lnkd.in/gxM7h9E9
🔸 See gradients flow in real-time: https://lnkd.in/gKdwgAzz

4️⃣ Build Real Networks
🔸 Aladdin's playlist (underrated gem): https://lnkd.in/gP3k_rNx
🔸 NLP from scratch (no shortcuts): https://lnkd.in/g77inpPa
🔸 Adversarial examples (break things): https://lnkd.in/gFDzkep8

5️⃣ Production PyTorch (Where 90% Fail)
🔸 Distributed training without tears: https://lnkd.in/gdPZDnxB
🔸 Profile before you scale: https://lnkd.in/gfs7JQ6R
🔸 Mixed precision (2x speedup): https://lnkd.in/gikyETYK
🔸 Performance tuning secrets: https://lnkd.in/g8BE7J28

The uncomfortable truth:

Most "PyTorch experts" can't debug a CUDA out-of-memory error.
They copy tutorials. They call high-level APIs.

But when production breaks? Crickets.

Start with micrograd. Build intuition.
Then scale up. One tensor at a time.

What PyTorch concept finally clicked for you?

♻️ Repost to help someone skip the 3am debugging sessions

The AI bubble is 75% of US GDP growth.

Read that again.

I spent the weekend analyzing Deutsche Bank's latest report.
What I found made me cancel my AI startup plans.

The dot-com bubble was 15.6% of GDP when it popped.
AI is already 5x bigger. And growing.

Here's the brutal reality nobody wants to admit:

1️⃣ Every AI Company Is Bleeding Cash
• OpenAI: Losing money on every $200/month subscription
• Break-even price: $2,000/month (10x current)
• 2026 projected loss: $14 billion
• 2029 projected loss: "Hundreds of billions"

2️⃣ AI Infrastructure Is a Money Pit
Data centers need $2 trillion annual revenue to break even.
Actual AI industry revenue: $390 billion.
That's a $1.6 trillion gap. Every. Single. Year.

3️⃣ AI Adoption Is Dead in the Water
• 95% of AI pilots fail (MIT study)
• AI coding tools make devs SLOWER (METR report)
• 77% of workers say AI increased their workload
• Adoption rates missed predictions 2 years running

The uncomfortable truth:

This isn't 1999. When the dot-com bubble popped, e-commerce was real.
It bounced back.

AI profitability is mathematically impossible.
The efficient compute frontier guarantees it.
Floridi's Conjecture proves agentic AI can't exist.

When this pops, there's no bouncing back.

What's keeping it alive? Your 401k.
70% of US venture capital is flowing into AI.
Half the S&P 500's growth is just tech companies building data centers.

Remove AI spending and US growth is already at zero.
We're IN the recession. AI is just hiding it.

♻️ Repost before your retirement fund discovers math

AI agent bootcamps charge $2,000 for this. You're getting it for free.

1️⃣ Design (6+ hours)
→ Langflow docs: docs.langflow.org
→ Travel planning agent template: https://lnkd.in/gmbWX5Mg
→ Agentic design patterns by Andrew Ng: 

2️⃣ Build (22+ hours)
→ LangGraph course: https://lnkd.in/g7raxjpC
→ LangSmith (modules 1-3): smith.langchain.com
→ Evaluation: https://lnkd.in/ghE-8EwA
→ Prompt engineering: 4 Methods of Prompt Engineering

3️⃣ Deploy (35+ hours)
→ LangChain Academy (modules 1-6): academy.langchain.com
→ MLOps specialization: https://lnkd.in/gmeb53UP
→ System design fundamentals: YouTube playlist
→ Distilabel for synthetic data: https://lnkd.in/gd_jJaJE

63 hours. $0.
The $2,000 bootcamp teaches the same stack with better marketing.

💾 Save this. It's the curriculum.
♻️ Repost to save someone $2,000.

Wanna build AI agents? Stop watching random tutorials.

Here's the actual learning path:

1️⃣ Design (6+ hours)
→ Langflow intro: langflow.org
→ Multi-agent workflow patterns
→ Build: Travel planning multi-agent system

2️⃣ Build (22+ hours)
→ LangGraph course: https://lnkd.in/g7raxjpC
→ LangSmith for tracing: smith.langchain.com
→ Evaluation metrics and prompt engineering

3️⃣ Deploy (35+ hours)
→ LangChain Academy: academy.langchain.com
→ MLOps fundamentals
→ LangGraph Cloud deployment

Total: 63 hours. If you follow the order.
6 months if you don't.

💾 Save this.  Start at Phase 1.
♻️ Repost if you've ever jumped straight to deployment and regretted it.

Your AI agent has dementia. LangChain just dropped the cure: 

DeepAgents CLI with actual persistent memory.

This project solves the most infuriating problem in AI development, agents that forget everything the moment you close your terminal. 

It gives your AI assistant an actual memory that persists across sessions. 

Key Features:

✅ True persistent memory
Your agent writes markdown notes to ~/.deepagents/memories/ and reads them before every response. Like a human taking notes.

✅ File operations that work: 
Read, write, and edit files in your project with diff previews and human approval. No more "I can't access your filesystem."

✅ Shell command execution
Run tests, install packages, deploy code - all with your permission. Because we're not insane.

✅ Multi-agent personalities
Create specialized agents (backend-dev, frontend-wizard, devops-guru) each with their own memory and expertise.

Getting started is stupidly simple. 
Just install and run: pip install deepagents-cli && deepagents

Harrison said they built this because copy-pasting context 47 times a day is insanity (we agree).

Check out the code, see the full toolset, and contribute on GitHub!

♻️ Repost if you're done being your AI's memory bank

90,000 AI papers published on Arxiv.

Ilya Sutskever says you only need 27 to understand 90% of what matters.

I just went through his list. Here's what shocked me:

What's NOT on the list:
❌ No RAG papers (everyone's obsessed with retrieval)
❌ No prompt engineering papers (despite the hype)
❌ No LangChain or agent frameworks
❌ Nothing about fine-tuning optimization
❌ Zero papers on "emergent abilities"

What IS on the list:
✅ "Attention is All You Need" (obviously)
✅ "The Unreasonable Effectiveness of RNNs" (2015, still matters)
✅ "Scaling Laws for Neural Language Models" (this is the one)
✅ "Kolmogorov Complexity and Algorithmic Randomness" (wait, what?)

The pattern I noticed:

1️⃣ Ilya's not telling you to read papers about using AI.
He's telling you to read papers about how intelligence works.
Three categories dominate his list:
Fundamentals (the architecture breakthroughs)
Attention mechanisms
Residual connections
LSTMs and sequence modeling

2️⃣ Scaling (why bigger actually works)
Scaling laws
Pipeline parallelism
Training dynamics

3️⃣ Compression (the theory underneath)
Minimum description length
Kolmogorov complexity
Information theory

Here's the controversial take:
If you're only reading papers from 2023-2024, you're studying tactics.

Ilya's list is strategy:

"Attention is All You Need" is from 2017.
"Understanding LSTM Networks" is from 2015.
"Deep Residual Learning" is from 2015.
The foundations haven't changed.
We're just scaling them.

The most surprising paper on the list:

"The First Law of Complexodynamics"
It's not even about neural networks.
It's about how complex systems evolve.

This is what separates engineers who use AI from engineers who understand AI.

If you only read 5:

1) Attention is All You Need (transformers)
2) Scaling Laws for Neural Language Models (why GPT works)
3) Understanding LSTM Networks (sequence modeling)
4) Deep Residual Learning (why deep networks train)
5) Kolmogorov Complexity (the theory of intelligence)

♻️ Repost to save someone from FOMO-reading 90,000 papers

There's a special hell in AI/ML engineering.

It hits around Year 2.
I call it the "valley of almost useful."

You're past "Hello World" tutorials.
Not good enough for real jobs.
Stuck building demos nobody uses.

The daily torture:

Morning: "Today I'll implement that paper"
Noon: Your embeddings are somehow 7-dimensional
Evening: Stack Overflow has zero answers
Night: "Maybe I should go back to web dev"

The worst part?

You know just enough to be dangerous.
Just not enough to be valuable.

🔥 The 3 Levels of "Almost Useful"

Level 1: The Overengineering Phase
Every problem needs a neural network.
CSV parsing? Better use GPT-4.
Your hello world has 17 dependencies.

Level 2: The Imposter Syndrome Pit
Everyone else seems to ship AI products daily.
Your Twitter feed is pure pain.
"Just deployed my 5th AI startup this week!"
Meanwhile, you can't get embeddings to work.

Level 3: The Clarity Moment
You realize 90% of "AI Engineers" are just prompt engineers.
Most production AI is boring Python scripts.
The magic isn't in the model - it's in the problem.

Here's what got me through the valley:

1️⃣ Ship broken things
My first production model had 67% accuracy.
Still saved the company $200K/year.
Perfect is the enemy of deployed.

2️⃣ Join a startup, not FAANG
Startups let you touch everything.
Break things. Learn fast. Ship faster.
Big tech will teach you one tiny slice.

3️⃣ Build boring solutions
Stop chasing cutting-edge.
Most businesses need basic classification.
Not your 12-layer transformer architecture.

4️⃣ Document everything
Future you will thank present you.
Especially when debugging at 3 AM.
"Why did I set embedding_dim=7?"

5️⃣ Find your valley friends
The loneliest part is thinking you're alone.
You're not. We all went through this.
Some of us just hide it better.

The uncomfortable truth:

The valley of almost useful is necessary.
It's where you learn that AI isn't magic.
It's engineering.

And engineering is mostly debugging.
And crying.
But mostly debugging.

Still in the valley?
Good. That's where real engineers are made.

Not in bootcamps.
Not in courses.
In the trenches, shipping broken code that somehow works.

Welcome to the club.

P.S. That senior engineer who seems to know everything? They googled it 5 minutes before the meeting. We're all just better at pretending now.

♻️ Repost to help someone realize they're not alone in the valley

Just watched an ML engineer at a fintech startup build a production fraud detection model in 2 hours.

From raw data to deployed.
No feature engineering. No pipeline code.
She just typed what she needed in plain English.

I was skeptical. Had to see it myself.

She opened Plexe, typed "fraud detection for card transactions" and pointed it at their Postgres db.
That's it.

2 hours later: deployed model with 0.89 AUC on 3M transactions.
Here's what actually happens: 

1️⃣ Describe your problem in plain English
"Product embedding model for search"
"Customer lifetime value predictor"
"Fraud detection system"
No templates. No configuration files.

2️⃣ Point it at your data
Postgres, Snowflake, S3, whatever you're using
It analyzes your schema automatically
No manual feature mapping

3️⃣ It handles what AutoML skips
• Engineers features from raw data (AutoML just uses columns as-is)
• Generates comprehensive evaluation reports with failure analysis
• Shows exactly where and why your model breaks
• Gives prioritized improvement recommendations
• Deploys with monitoring

The critical difference:
❌ AutoML: "Here's a model with 0.87 accuracy. Good luck."
✅ Plexe: "Here's why it works on these patterns, why it fails on those edge cases, and the three changes that will improve it most."

This is what changes:
You stop being a data plumber.
You start being a problem solver again.

Best part? 

Open-source core (Apache 2.0): https://lnkd.in/gEx42Q9b
Try it with $20 free credits: https://console.plexe.ai
Use code: LAUNCHDAY20

What ML model have you been putting off because wrestling with pandas makes you want to scream?

♻️ Repost to save someone 3 months of pipeline work

I pushed every agent framework until it broke. Most broke fast.

CrewAI vs LangGraph vs Agno vs PydanticAI vs Atomic Agents...
Plus SmolAgents, Mastra, and AutoGen.

The brutal truth:

Most devs hate the popular ones.

CrewAI and AutoGen? "Tricky to debug"
LangGraph? "Overly complicated abstractions"
SmolAgents' CodeAgent? "Cool idea, doesn't always work"

What actually matters:

🎯 Control vs Agency
Some frameworks let LLMs run wild (AutoGen)
Others keep them on a tight leash (Atomic Agents)
Guess which ones actually ship to production?

🛠️ Abstraction Level
High abstraction = Quick start, debugging nightmare
Low abstraction = More code, but you know what broke
Pick your poison.

💀 The Learning Curve
LangGraph: Steep as hell, but scales
CrewAI: Easy start, hits a wall
PydanticAI: Write everything yourself

The uncomfortable truth:
"AI agent frameworks are some of the worst forms of abstraction"

But here's what actually works:
• Agno for developer experience
• LangGraph for multi-agent at scale
• PydanticAI when you need control
• CrewAI if you're just prototyping

Half these frameworks won't exist in 2 years.
The winners? Whoever makes debugging possible.

Saved you 2 weeks of testing. You're welcome.
♻️ Repost if you've been burned by agent framework promises

a16z just published their AI coding report.

The math blew my mind:

AI coding costs $10k/year per developer.
But creates $100k+ in value.
That's $3 trillion in new GDP.

The entire economy of France. From a few Silicon Valley startups.

Here's what nobody's talking about:

Software development just shifted from pure people cost to people + compute cost.

Your junior developer in Bangalore?
Now competing with Claude at $10k/year.

The uncomfortable truth:

We're watching the end of cost-based outsourcing.

Not because AI replaces developers.
But because location doesn't matter when compute costs dominate.

I tested the numbers myself:

• 100k context window filled
• Claude Opus 4.1 reasoning mode
• 10k output tokens
• 3 queries/hour, 7 hours/day

= $10,500 annually

That's more than many offshore developers.

But here's the kicker:

The companies winning aren't cutting developers.
They're hiring MORE.

Why?

Because when productivity doubles, the math changes:
• 1 developer + AI = 2x output
• Cost per feature drops 50%
• Suddenly every project has positive ROI

The report shows clear winners already:

Cursor: $500M ARR in 15 months
Windsurf: $2.4B Google acquisition
Claude Code: Started a war with its own distribution channel

The pattern is obvious:

Lower barriers → More projects → More complexity → More developers needed

AWS didn't kill sysadmins.
WordPress didn't kill web developers.
AI won't kill software engineers.

It just changes what they do.

University CS programs?
Already obsolete.

Teaching algorithms when AI writes them faster?
Teaching syntax when AI handles implementation?
Teaching debugging when AI prevents most bugs?

That's like teaching calligraphy in the age of keyboards.

The real skill now:

Knowing when the AI is hallucinating.
Dragging it out of architectural holes.
Translating business needs into prompts.

The $3 trillion opportunity isn't in replacing developers.

It's in making every developer 10x more valuable.

While universities update their curriculums,
smart developers are already adapting.

Are you?

♻️ Repost if you think CS education needs a complete overhaul

Plexe is a Python library that turns "I need to predict customer churn" into a working ML model.

No feature engineering. No pipeline code. Just natural language intent and your data.

It integrates with OpenAI, Anthropic, and Ray, making it easier to build sentiment analysis, fraud detection, and forecasting models in minutes instead of weeks.

Plexe doesn't just build the model - it generates a thorough evaluation report showing exactly where and why it performs well (or doesn't).

A powerful tool for anyone working at the intersection of product development and machine learning.

♻️ Repost if you've ever deployed a model and just hoped it worked.

🎁 Launch special: Get $20 free credits → https://console.plexe.ai
Code: LAUNCHDAY20

The most underrated AI agents video on YouTube has 155K views.

It should have 15 millions.

Dexter breaks down why coding agents actually fail in production.
Not theory. What broke when they shipped real agent systems to real software teams.

The line that changed how I think about agents:
"Everything is context engineering."

He shows:

→ Why naive back-and-forth prompting doesn't scale (and what does)
→ Spec-first development that keeps agents on track
→ Context compaction strategies for long-running workflows
→ How subagents handle complexity without exploding

No demo theater. Just hard-won lessons from production.

If you're building agents that need to work beyond the prototype, this is required watching.

Watch it: https://lnkd.in/e99smz5g

♻️ Repost if you've hit the "context limit exceeded" wall

I just watched Sebastian Raschka build the entire attention mechanism from scratch in 2 hours.

No abstractions. No "just use torch.nn.MultiHeadAttention."

Line-by-line Python until it clicks.

This is 's breakdown of how LLMs actually work.

Step 1️⃣ Basic self-attention (no weights) 
Shows you the raw math: how a model decides which words matter.
You see why "bank" in "river bank" gets different attention than "bank" in "savings bank."

Step 2️⃣ Add learnable parameters 
Now the model can learn what to pay attention to.
This is where it stops being a simple lookup and becomes intelligence.

Step 3️⃣ Causal masking 
The trick that makes GPT work: tokens can only attend to previous tokens.
Without this, the model cheats during training.

Step 4️⃣ Multi-head attention 
The breakthrough: run 8 attention mechanisms in parallel.
One head learns syntax. Another learns semantics. Another learns long-range dependencies.
This is why transformers work.

👉 What makes this different:

Most tutorials either:
Show you the math (incomprehensible)
Show you the PyTorch API (you learn nothing)

Sebastian does both.

You see the math implemented in code, step by step.

By the end, you understand why attention_weights = softmax(Q @ K.T / sqrt(d_k)) isn't magic.

It's just weighted averaging with learnable weights.

💡The moment it clicked for me:

Attention isn't complicated.
It's just "which words should influence this word's representation?"
Then you do that for every word.
Then you stack it 96 times (GPT-4).

That's it.

The Lovable for Data Analysis Just Dropped

Upload a file → Ask a question → Get your answer.

No Python, no SQL, no notebooks required.

Compare that to what you’re doing now: 

❎ Write SQL queries or Python scripts 
❎ Debug when they break
❎ Switch between notebooks, terminals, dashboards 
❎ Wait for a data scientist to help you

This agent actually understands your data:

✅ Sees your structure, knows how to query it 
✅ Maintains context across your workflow 
✅ Fixes its own errors
✅ Generates insights, not code you need to interpret

This is a big deal for anyone doing real data analysis work. 

You get the power of a data science team without needing to become one.

The best part?

I've got $15 in free credits for you to test it (no expiration date).

♻️ Repost + comment "GIVE ME MY WEEKENDS BACK 😭" and I'll send you the link

Traditional RAG: retrieve → generate → pray

Agentic RAG: retrieve → validate → retry → succeed

One is a pipeline. The other is a feedback loop.

I just built this Agentic RAG system with LangGraph + Milvus

It knows when it's wrong, and fixes itself

Here’s how:

Router: "Can I answer without retrieval?" 
✅ Yes: Skip to generation 
❌ No: Retrieve from Milvus

Grader: "Are these chunks actually relevant?" 
✅ Good: Generate answer 
❌ Bad: Rewrite query, loop back

Validator: "Does this answer match the sources?" 
 ✅ Yes: Return to user 
❌ No: Retry with different approach

It's a graph with feedback loops, not a one-shot pipeline.

The difference:
Traditional RAG commits to one path. No second chances.
Agentic RAG loops back, validates, catches mistakes, and retries until it's right.

You get a system that debugs itself before you see the output.

Why this stack:
LangGraph = control flow primitives (routing, looping, branching) 
Milvus, created by Zilliz = production-grade vector search at scale
Together: RAG that thinks instead of guesses.

Full implementation 👇

This is the RAG architecture pattern that will become standard.
♻️ Repost to save a fellow engineer from another week of RAG debugging

BFCM isn't won by the brand with the best products.

It's won by the brand that sees what's breaking first.

Most operators spend BFCM like this: 

9 AM: Campaign's crushing it, scale to $10K/day
11 AM: ROAS drops 40%, nobody knows why
2 PM: Still diagnosing
5 PM: $6K burned on a broken campaign

The strategy wasn't wrong. 

You just couldn't see the problem fast enough. 

Diagnosing "why is ROAS dropping?" right now: 

→ Open Meta Ads Manager
→ Check Shopify analytics
→ Pull GA4
→ Cross-reference Klaviyo
→ Export to spreadsheet
→ Try to connect the dots

20 minutes later you've burned $3K.

Triple Whale + Moby give you real-time answers from your complete data stack. 

👉 https://lnkd.in/gtgB8GJw

Moby has real-time access to all of it. 

Ask Moby: "Why did ROAS drop?"

Get: "iOS campaign CTR crashed 67% after creative refresh. Top SKU went out of stock an hour ago. Traffic's converting to lower-margin backup products."

Not a guess. The actual answer.

One brand caught this: 

The campaign was showing 187% ROAS. 
Moby a flagged 41% return rate. 
Actual ROAS after was returns: 73%. 

They were about to scale into a loss.

During BFCM, 20 minutes to diagnose = thousands in wasted spend. 

The brands that win can see problems in seconds, not hours. 

And yes, I am a Triple Whale partner.

👉 https://lnkd.in/gtgB8GJw

♻️ Repost if you've burned budget figuring out what went wrong

taplio

The Battle

Alex Razvant

Paolo Perrone

Followers

Posts

Engagements

Likes

Performance ScoreAlex Razvant's LinkedIn Performance

Performance ScorePaolo Perrone's LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Alex Razvant

Paolo Perrone

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);}Alex Razvant's LinkedIn Performance

Performance ScorePaolo Perrone's LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreAlex Razvant's LinkedIn Performance