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.

Carbaba AI

• Founded Carbaba Cloud Service, a pioneering B2B SaaS company in the UAE focused on compliance systems and QA validators.  
• Spearheaded the transformation of document intelligence, enhancing documentation workflows for enterprises.  
• Developed innovative solutions that improved operational efficiency, resulting in a 40% reduction in document processing time.

General Manager

CliniNote

CliniNote is a rapidly growing startup making waves in the healthcare industry. CliniNote develops innovative tools for real-world data collection and assistant software specifically designed for doctors, hospitals,pharmaceutical companies, and clinical trials, with a global reach. The team, along with their business angels and investors, share a strong commitment to integrity, consistency, and focus. They’re passionate about driving positive change in healthcare through the power of digital solutions and
real-world data, ultimately saving lives.

Software Developer

Monash University

Engaged in mentoring students embarking on their journeys in data analytics, providing a bridge
between academic learning and real-world applications.

Delivered project-based feedback and career advice, drawing from my extensive experience in
data science and biomedical engineering.

Education Mentor

Swinburne University of Technology

Working on research project to build a new systematic artificial intelligence program using intricate machine learning classification and regression algorithms to predict heart disease using dataset from UCI repository. 
These techniques require the adaptation of high level feature engineering strategies to acquire an accurate smart medical technology. 
This new technology can interpret detailed analogies of patients to create a sustainable, normalized algorithm that predicts different types of heart disease directly proportional to the valves diameter narrowing. 
This technology utilizes a combination of R programming language to visualize and preprocess the dataset accordingly and python language to utilize the agility and robustness for an enhanced classification technology.

Machine Learning Researcher

Fiverr

Dealing with segregating data for different online businesses 
Cleaning repetitive data 
Inputting data into database 
Having to deal with images and string format
High level focus on detail

Data Entry Associate

- Teaching statistical methodologies to RMIT university students studying Statistics & Data Analytics

- Helped RMIT university students with statistics and marketing-related assignments 

- Worked with my students to plan and construct a strategy to tackle difficult projects that required Data analytics and statistical expertise. These involved but not limited to: 
- Ideation stages of understanding the problem and planning a timeline for each member of the group to perform their roles accordingly. 
- Processing stage where the group started showing improvements on their allocated tasks. This stage heavily utilized proficient MS Excel expertise. 
- Finalizing stage which involves organizing the ideas together in a compact format to present the project in a visually pleasing and effective manner to the client.

Private Tutor

CarbabaAI, a UAE-based B2B SaaS company, pioneers document intelligence using AI-driven compliance systems and QA validators to optimize enterprise workflows. As Founder, contributions include innovative solutions that reduced document processing time by 40%, showcasing a commitment to operational efficiency and transformative technologies.  

With a biomedical degree major with a data science minor from Swinburne University of Technology, expertise spans machine learning, AI systems, and data analytics. Previous roles, such as Software Developer at CliniNote, focused on real-world data solutions for healthcare. Additionally, mentoring at Monash University bridged academic concepts with practical applications, reflecting a dedication to responsible AI-driven industry innovation.

Ghadeer  A.

Why 90% of ML Engineers Quit Before Year 2 (SAVE THIS 👇)

It doesn’t have to be this way

Most learners drown in models with no map.
Nothing sticks.

Concepts blur together.
Progress stalls fast.

A full visual system fixes everything.
Clarity arrives instantly.  

✅ This collection maps Neural Networks end-to-end.
Page 1 shows every major architecture in one view.
You see how each model connects across the ecosystem.

✅ The Neural Network Graphs on page 2 break down flow.
They show feed-forward, recurrent, LSTM and GRU paths.
Patterns become visible instead of abstract.

✅ Pages 3–4 decode the maths behind learning.
Not theory for theory’s sake. Practical intuition.
You finally understand why models behave as they do.

✅ Machine Learning in Emoji on page 5 simplifies concepts.
Complex ideas become instantly memorable.
Supervised, unsupervised and reinforcement made visual.

✅ Scikit-Learn, Azure ML and Python basics span pages 6–10.
They show algorithms, workflows and starter code together.
Zero guesswork on what to use and when.

✅ Pandas, NumPy and data wrangling fill pages 11–15.
You see cleaning, shaping, merging and plotting clearly.
Data stops being messy and starts being manageable.

✅ SciPy, Matplotlib and ggplot2 visuals on pages 16–18.
They reveal analysis, plotting and scale choices.
Your data storytelling sharpens instantly.

✅ PySpark and Big-O on pages 19–20 zoom out further.
You understand scalability and operations at a glance.
It becomes the roadmap many engineers never get.

This is the satellite view.
This is the missing map.
This is how AI finally clicks.
This is how you build intuition fast.

➕ Follow Ghadeer for more insights
♻️ Repost to help others in your network
📩 Save for later

I Help AEC Firms Automate Their QA Processes | Founder @ Carbaba | AI-Native Document Intelligence

Neural_Networks_Cheatsheet

95% of “AI agents” aren’t agents. 
They’re prompt wrappers with dreams.

Real autonomy isn’t magic.
It’s layers. And almost everyone is missing three of them.

Here’s the stack the top 3% of builders actually use: 👇

🔹Layer 1: LLMs — The Raw Intelligence

Great at language. Useless at autonomy.
	•	Tokenization + inference
	•	Prompt shaping
	•	Direct API access

This is where most teams stop — then wonder why nothing is reliable.

🔹Layer 2: AI Agents — Intelligence That Acts

A brain finally gets hands.
	•	Tools + function calling
	•	CoT / ReAct reasoning
	•	Task decomposition
	•	Long-term memory

One agent is smart. But one agent alone can’t run a real workflow.


🔹Layer 3: Agentic Systems — Intelligence That Cooperates

This is where autonomy becomes repeatable.
	•	Multi-agent communication
	•	Roles + specialization
	•	Routing + scheduling
	•	Multi-agent RAG for shared context

This is the leap from “bot” → “operation.”

🔹Layer 4: Agentic Infrastructure — Intelligence That Survives Reality

The layer everyone forgets… until their system fails.
	•	Observability + traceability
	•	Error handling + retries
	•	Security + compliance
	•	Human-in-the-loop control
	•	Workflow automation

Without this layer?
Your agents are demos.
With it?
Your agents become deployable systems.

Most teams think they’re building agents.
They’re not.
They’re stacking prompts and praying.

Real autonomy lives at Layer 4.
Real reliability comes from architecture, not clever prompting.

➕ Follow Ghadeer for more agent frameworks
♻ Repost to break the “one-layer agent” myth
📩 Save this for your next system design session

Ghadeer  Al Ruwaishedi

J.P. Morgan Can’t Deploy AI Agents That Follow ‘Vibes.’ Here’s the Framework Built for Reality (SAVE this 👇)

Reliability rules.

Agents impress in demos.
Then collapse when real users touch them.
The failure is structural.

Your agents struggle.
Not from prompts.
From fragile frameworks.

Most systems were never designed for:
consistency
compliance
explainability
rule enforcement
auditability
controlled tool use

Parlant was.

It removes randomness entirely.
Through five engineered components.

👉 Journeys
Structured flows that replace prompt luck.
Predictable steps that keep agents stable.

👉 Guidelines
Enforced tone, behavior, safety rules.
Your musts and nevers hard-wired.

👉 Relationships
Dependencies and edge-case logic.
The reasoning layer others ignore.

👉 Tools and Retrievers
Scoped, governed access to APIs and data.
No unauthorized calls. No surprises.

👉 ARQs
Invisible audits on every decision.
Automatic checks before output leaves the agent.

Other frameworks teach responses.
Parlant teaches reliability.
And reliability is the new intelligence.

If your agents touch onboarding, QA/QC, compliance, support, routing, or enterprise workflows,
Parlant ensures they never go off-script.

Build agents that follow rules.
Build agents that survive reality.
Build agents you can trust to scale.

Check out Parlant’s repo: https://lnkd.in/dywtVkwu

➕ Follow Ghadeer for more insights
♻️ Repost to help others in your network
📩 Save for later

No alternative text description for this image

Stop. Building. Agents. Without. Protocols. You’re literally wasting money. 
Here’s why—

Protocols print power.

Everyone’s chasing better models.
Faster inference. Smarter reasoning.
But without protocols agents can’t talk.

They stall.
They duplicate effort.
They burn compute, time, and budget.

Here’s what’s fixing that (and why it matters) 👇:

1️⃣ MCP (Model Context Protocol) — Anthropic
The USB-C of AI agents.
Connects LLMs securely to real-world tools.
Two-way context, zero friction.

2️⃣ A2A (Agent 2 Agent) — Google
Agents finally speak the same language.
They can discover, collaborate, and execute across vendors.
Goodbye silos, hello synergy.

3️⃣ ACP (Agent Communication Protocol) — IBM
Think REST, but for autonomous agents.
Persistent, asynchronous, and scalable.
Built for observability, not chaos.

4️⃣ SLIM (Secure Low-Latency Interactive Messaging) — Cisco
Real-time, pub/sub communication over gRPC.
Perfect for event-driven AI systems.
Fast. Encrypted. Reliable.

5️⃣ Agora — Oxford University
Turns plain language into live protocols.
Dynamic, user-centric workflows on demand.
Where natural speech becomes structured action.

6️⃣ ANP (Agent Network Protocol) — ANP Team
Powered by W3C DIDs and JSON-LD.
Enables decentralized agent identity and trust.
The backbone of interoperable ecosystems.

7️⃣ AG-UI (Agent-User Interaction) — CopilotKit
The front-end bridge between users and agents.
Standardizes how interfaces and AI collaborate.
Seamless human-agent symbiosis.

AI models are getting stronger.
But without protocols, your agents are shouting into the void.
Protocols aren’t “nice-to-haves.” They’re infrastructure.
They’re the difference between chaos and scale.

Link to repos 👇

➕ Follow Ghadeer for more insights
♻️ Repost to help others in your network
📩 Save for later

7_Protocols_used_in_AI_Agents

🚨 90% of teams building AI agents are doing it wrong. Google just confirmed it.

They’re treating agents like glorified workflows:

Prompt → Tool call → Result → Ship.

That’s why their “agents” fall apart with edge cases, hallucinate under pressure, and behave like unpaid interns.

After studying Google’s new Agent Architecture playbook, I realized the real game isn’t tool use.
It’s orchestration and context engineering.

Agents don’t execute tasks they curate context, plan, act, observe, and iterate.

My workflow now:

1️⃣ Start with the agent’s mission, not the prompt
2️⃣ Map the Think → Act → Observe loop
3️⃣ Define tools as capabilities, not “API wrappers”
4️⃣ Add memory + policy + identity
5️⃣ Stress-test the orchestration layer, not the model

This isn’t automation.
This is systems design.
The best engineers aren’t prompting they’re orchestrating autonomous problem-solvers.

We’re moving from “LLMs in a wrapper” to full agentic operating systems with planning, memory, delegation, governance, and self-improvement.

The billion-dollar gap?
Tools that help teams build orchestration layers, not prompts.

We’re not being replaced.
We’re being promoted into agent architects.

➕ Follow Ghadeer for more insights
♻️ Repost to help others in your network
📩 Save for later

Introduction_to_Agents

If you’ve ever spent 3am debugging why your PDF parser thinks a table is 47 separate paragraphs, this one’s for you

OCR grew up.

For years, OCR meant “text out, chaos in.”
Tables shattered. Layouts lost. Context gone.

Every document looked fine until your LLM choked on spaghetti text.
Every fix meant regexes, retries, and rage.

In 2025, OCR finally became document intelligence.

6 systems now read like humans not scanners. 👇

1️⃣ Google Cloud Document AI
Reads scanned and digital PDFs in one pipeline.
Detects tables, key values, even handwriting.
Perfect for financial and education documents.

2️⃣ Amazon Textract
Sync for small, async for scale.
Pulls text, forms, tables, signatures — clean JSON out.
Ideal for AWS-native IDP and claim pipelines.

3️⃣ Microsoft Azure Document Intelligence
Custom and prebuilt models now run in containers.
Extracts structure, figures, and layout for LLM input.
Hybrid ready for air-gapped deployments.

4️⃣ ABBYY FineReader & FlexiCapture
Still the on-prem heavyweight.
190+ languages, tight compliance, deep zoning.
Flawless for regulated or multilingual workloads.

5️⃣ PaddleOCR 3.0
Free. Open source. Apache licensed.
Covers detection, recognition, and structure.
Built for startups building self-hosted RAG pipelines.

6️⃣ DeepSeek OCR
A new species: compression-first OCR.
Encodes long documents into vision-language form.
10x smaller input for LLMs, open-source, GPU-ready.

OCR used to see letters.
Now it understands meaning.
From extraction to interpretation that’s the leap.

For AI teams, this means fewer hacks, faster pipelines,
and context that finally survives the parser.

OCR taught us patience.
And creative regex therapy.
But now, it finally works like we always wanted it to.

➕ Follow Ghadeer for more insights
♻️ Repost to help others in your network
📩 Save for later

6_Top_OCR_Models

I’m building a $1M company with $0 employees. Here’s the tech stack replacing my entire team 👇

Humans are optional. Systems aren’t.

Most founders burn out doing everything.
I built with stack that does everything for me.
Because efficiency beats effort. EVERY time.

Too many founders trade time for growth.
They build a business that runs on stress.
I built one that runs on systems.

Cursor writes my code faster than any dev.
Perplexity and Gemini handle deep market research.
ChatGPT, Grox, and Claude write copy that converts.
Midjourney, Recraft, Figma, and Canva bring visuals to life.
RecCloud, Tidio, Zapier, Calendly, and PicWish automate the grind.

This isn’t a team. It’s a machine.
And it never asks for a vacation.
You don’t need more people.
You need better tools.

➕ Follow Ghadeer for more insights
♻️ Repost to help others in your network
📩 Save for later

li_at

JSESSIONID

If you want to build AI agents that run autonomously in production (not just demos), master these 15 skills:

1. Python Programming: Beyond Basic Scripts
Everything in Agentic AI runs on Python.
Learn async programming, error handling, and class design.

2. Prompt Engineering as Code
Stop guessing what to write.
Structure prompts with templates, variables, and version control.

3. LLM Selection & Token Economics
Understand context windows, streaming, and cost per call.
Pick the right model for each task (GPT-4 isn’t always the answer).

4. API Integration & Webhook Design
Agents can’t act alone.
Connect them to Slack, Gmail, databases, and custom endpoints.

5. Workflow Automation Architecture
Chain triggers, conditions, and actions that scale.
Master tools like Zapier, n8n, and Make

6. JSON Schema & Data Validation
Agents speak JSON.
Design schemas that prevent hallucinated outputs and data chaos.

7. Vector Databases & Semantic Search
Give your agents persistent memory.
Weaviate, Pinecone, and Chroma store embeddings for context retrieval.

8. Retrieval-Augmented Generation (RAG)
Let agents access external knowledge bases.
Build pipelines that inject relevant context into every response.

9. Containerization & Cloud Deployment
Run agents 24/7 without your laptop.
Docker, AWS Lambda, and serverless functions keep them alive.

10. Agent Frameworks That Actually Scale
LangChain, CrewAI, and AutoGen aren’t just libraries.
They’re architectures for building multi-step autonomous systems.

11. Tool-Use & Function Calling
Agents that use calculators, search engines, and APIs are 10x more powerful.
Teach them when and how to reach for external tools.

12. Multi-Agent Orchestration
One agent is a script. Multiple agents is a workforce.
Design systems where specialists collaborate on complex tasks.

13. Memory Management: Short-Term vs Long-Term
Stateless agents forget everything.
Build conversation history, user preferences, and learned patterns.

14. Observability: Logging, Tracing & Debugging
Production agents fail silently.
Use LangSmith, Sentry, and custom logs to catch issues before users do.

15. Security, Rate Limits & Ethical Guardrails
Agents without constraints are liabilities.
Implement permissions, content filters, and spend caps.

The AI race isn’t about prompts anymore.

It’s about building autonomous systems that think, act, and learn while you sleep.

Companies don’t pay for demos. They pay for agents that handle real workflows in production.

These 15 skills separate builders from prompters.

Which skill are you struggling with most? 👇

➕ Follow Ghadeer A.  for more insights
♻️ Repost to help others in your network
📩 Save for later

Retrieval = intern. Reasoning = analyst. Action = executive. Most companies are paying executive prices for intern output.

Most enterprise AI in mobility stops at retrieval.
It answers but it doesn’t understand.

At Carbaba, we go further with Agentic RAG.

Legacy RAG retrieves data.
But it lacks reasoning and adaptability.

That’s why fleet systems still need humans
to interpret, plan, and decide.

Agentic RAG merges retrieval, reasoning, and real-time action 
turning static AI into a dynamic decision-maker.

📌 RAG (Retrieval Augmented Generation)
Pulls precise vehicle, route, and system data.
Merges it with prompts for analysis.
Generates answers but without context awareness.

📌 AI Agents
Plan and reason across connected data sources.
Access live telematics, maintenance logs, and supply chains.
Deliver decisions not just insights.

📌 Agentic RAG
Combines retrieval with intelligent action.
Agents select data sources, plan retrieval paths,
and feed reasoning-driven input to the LLM.

🗄️ Operational Workflow
User query → routed to agent.
Agent plans → retrieves real-time data.
Context retained → output optimized for reasoning.
Response delivered → insight becomes autonomous action.

This is how Carbaba automates enterprise mobility 
from predictive maintenance to intelligent dispatch.

Agentic RAG isn’t the future. It’s the framework.
The enterprises building with it today
will define tomorrow’s intelligent systems.

➕ Follow Ghadeer for more insights
♻️ Repost to help others in your network
📩 Save for later

Google DeepMind’s Latest Move Just Made ‘Deep Learning’ Obsolete (Check It Out 👇)

The era of stacking layers is over.

Levels beat layers.

We didn’t lose the plot.
We were given the wrong plot.

Deep learning wasn’t deep.
It was just tall.

The pain we all know

More layers.
More FLOPs.
More cost.
Less learning.

Scaling became a blunt instrument.
Smarter models never arrived.
Just bigger bills.

Because depth never created intelligence.
It only created noise.

The reveal that changes everything

Nested learning exposes the truth:

LLMs don’t learn from data.
They learn from compressing their own context.

Attention = fast memory
MLPs = slow storage
Optimizers = learned recall

Each subsystem runs its own learning loop.
Different speed.
Different purpose.
Different update rhythm.

We didn’t build deep nets.
We built multi-timescale minds.

Why this matters now

The next frontier isn’t bigger models.
It’s multi-level learning systems.

More levels =
⚫️ stronger in-context learning
⚫️ longer retention
⚫️ continual adaptation
⚫️ self-modifying behavior

HOPE proves it:
A model rewriting its own update algorithm during inference.
Not training.
Not fine-tuning.
Evolving in real time.

The identity flip

You’re not training models anymore.
You’re engineering learning hierarchies.

Memory depths.
Update rhythms.
Nested objectives.
Multi-level gradients.

Welcome to post–deep learning.

➕ Follow Ghadeer for more insights
♻️ Repost to help others in your network
📩 Save for later

Nested_Learning

Every paid OCR tool failed on my 700-page document. So I built a free pipeline that reads like a human.

Last month, one document nearly broke me.

A 700-page scanned tender (charts, tables, signatures, handwritten notes) completely shattered every paid OCR tool I tried.

Too slow.
Too messy.
Too inaccurate.
Too blind to structure.

The new generation of open OCR models isn’t behaving like traditional tools.
They’re interpreting layout, context, visuals, and semantics the way humans do.

Here are the 8 models that actually survived real enterprise documents 👇:

1️⃣ Nanonets-OCR2-3B

Handwriting, signatures, watermarks, flowcharts.
Semantic Markdown output. Multilingual.

2️⃣ PaddleOCR-VL

Old scans → clean HTML.
Tables + charts extracted with structure.
109 languages.

3️⃣ dots.ocr

Layout-grounded.
Understands text placement + images.
Natively multilingual.

4️⃣ OlmOCR-2

Enterprise-scale batches.
Consistent Markdown + LaTeX.
English-focused and reliable.

5️⃣ Granite-Docling-258M

Tiny, promptable, DocTags-friendly.
Fast + flexible for production.

6️⃣ DeepSeek-OCR

Ridiculously fast for large workloads.
Optimized for throughput + consistency.

7️⃣ Chandra

Grounded OCR for complex docs.
Great for technical + academic materials.

8️⃣ Qwen3-VL

Understands ancient scripts → modern layouts.
A true multimodal foundation.

Open models are finally outperforming traditional OCR 
faster, cheaper, smarter.
Not just reading… reasoning.

If you deal with tenders, compliance docs, audits, or enterprise workflows:

This is the Document AI shift you can’t ignore.

What’s the OCR issue that frustrates you the most?

➕ Follow Ghadeer 
♻ Repost to help others
📩 Save for later

OCR_Open_Models

I shipped 6 AI chatbots in 2024. Deleted all of them. This is what I built instead 👇

Chatbots are shallow.

Teams build demos.
Demos collapse fast.

Edge cases kill them.
Reality exposes gaps.

Agentic systems survive.

👉Building Agentic AI Systems by Anjanava Biswas & Wrick Talukdar  
Shows multi-agent orchestration that stays stable.
Reveals how agents reason, plan, and delegate.
Goes far beyond simple LLM wrappers.

Explains how to fuse GenAI, APIs, and knowledge graphs.
Keeps automation factual under messy conditions.
Shifted how I design systems that endure enterprise chaos.

👉AI Agents in Practice Valentina Alto  
Turns blueprint into production-ready execution.
Covers LangGraph orchestrators that don’t break.
Teaches memory and context management at scale.

Adds secure tool use and real error recovery.
Provides auditability for regulated workflows.
A playbook for ops, sales, compliance, and support.

🔥Why this mattered
Let me ship AI that handles edge cases at scale.
Self-corrects without oversight.
Coordinates tools with precision.
Survives reality instead of demo conditions.

Agentic AI replaces fragile demos.
It becomes operational infrastructure.
It delivers impact teams can trust.

Check out Packt’s publishing for more 👇 

➕ Follow Ghadeer for more insights
♻️ Repost to help others in your network
📩 Save for later

I just watched a non-technical person build a custom AI tool in 15 minutes. Game over (SAVE THIS👇)

The no-code era just leveled up.

Building AI tools used to be for engineers.
Now anyone can do it faster than a coffee break.
Claude Skills makes it effortless.

1️⃣ Turn on Claude Skills in settings.
2️⃣ Enable “skill-creator.” It builds other skills for you.
3️⃣ Ask Claude to create your custom skill. IT writes the file.
4️⃣ Download, upload, and test it by name.

That’s it. You’ve just automated your workflow.
No prompts. No Python. No problem.

Start simple. Let Claude handle the rest.
15 minutes. Infinite leverage.

➕ Follow Ghadeer for more insights
♻️ Repost to help others in your network
📩 Save for later

Why Most AI Agents Are Actually Broken
(And It Has Nothing to Do With the Model 👇)

Memory is the real failure point.

Agents forget.
Contexts drift.
Outputs collapse.

The problem isn’t intelligence.
It’s missing architecture.

Your agent loses track instantly.
You keep scaling the model to compensate.

Fix memory layers.
Not parameter counts.

Vector Memory

🔹 Plain vector RAG
Embedding lookups
Fast, simple, widely supported

Loses temporal structure.
Drifts semantically.
Forgets user decisions.

🔹 Tiered vector memory
Working set + archive

Reuses key info better.
Still suffers paging errors.
Creates agent divergence.

Graph Memory

🔹 Temporal KG memory
Cross-session reasoning
Strong temporal awareness

Requires schemas and pipelines.
Edges can go stale.

🔹 Knowledge-graph RAG
Hierarchical communities
Multi-doc, multi-hop logic

Heavy graph construction.
Bias and traceability overhead.

Event / Log Memory

🔹 Execution logs + checkpoints
Ordered versioned truth
Supports replay + repair

Prone to log bloat.
Instrumentation gaps matter.

🔹 Episodic long-term memory
Episodes + metadata
Long-horizon recall

Boundary errors occur.
Cross-agent misalignment appears.

Reliable agents need layers.
Vector for semantics.
Graph for structure.
Logs for truth.

That’s the line between toy agents
and production systems that don’t break.

➕ Follow Ghadeer for more insights
♻️ Repost to help others in your network
📩 Save for later

🚨 Most people “study AI” the wrong way they binge random GitHub repos.

The top 1% don’t do that.
They follow levels and each level unlocks the next.

Here’s the 3-tier AI learning ladder built from 12 elite repos (bookmark this 🔥).

LEVEL 1 — BEGINNER (Become AI-literate fast)

If you feel lost, these repos turn chaos into clarity.

1. Microsoft – AI Agents for Beginners
↳ https://lnkd.in/dyWxyqpR
The cleanest way to understand agents without drowning in jargon.

2. Machine Learning for Beginners (Microsoft)
↳ https://lnkd.in/dBj3BAEY
40 lessons, project-based, human-friendly. Zero fluff.

3. Prompt Engineering Guide
↳ https://lnkd.in/gJjGbxQr
The playbook for talking to models like a power-user, not a tourist.

4. LLM Course (Maxime Labonne)
↳ https://lnkd.in/d-_HXFSy
LLMs explained without assumptions. You’ll finally “get it.”

LEVEL 2 — INTERMEDIATE (Go from learner → practitioner)

This is where you stop consuming and start building.

5. GenAI Agents (Nir Diamant)
↳ https://lnkd.in/dxA9UQaJ
Real agent workflows, real patterns, real projects.

6. GenAI Agents 
↳ https://lnkd.in/dEt72MEy
Complementary examples + deeper architecture notes.

7. Made With ML
↳ https://lnkd.in/d2dMACMj
Production-grade ML thinking… the part 90% of people skip.

8. Hands-On Large Language Models
↳ https://lnkd.in/dxaVF86w
Build RAG, fine-tune, deploy—everything companies actually pay for.

9. Hands-On AI Engineering
↳ https://lnkd.in/dxgjH545
The missing link between “learning AI” and “working in AI.”

LEVEL 3 — EXPERT (For people aiming at $300K+ AI roles)

You don’t just learn AI here you start thinking like an AI architect.

10. Designing Machine Learning Systems
↳ https://lnkd.in/dEx8sQJK
The bible of scalable ML systems. Required reading for senior roles.

11. Awesome Generative AI Guide
↳ https://lnkd.in/dJ8gxp3a
A curated knowledge graph of everything GenAI. You’ll never be behind again.

12. The Ultimate LLM Engineering Path
(from combining the repos above)
↳ RAG → agents → evals → observability → scaling
At this point, you’re not “using AI.” You’re engineering ecosystems.

The truth nobody says out loud:

AI isn’t about learning everything.
It’s about learning the right things in the right sequence.

Most people stay stuck because they consume randomly.
But the people who rise fast follow levels not noise.

Which level are you currently on?
Drop it below 👇

➕ Follow Ghadeer for more insights
♻️ Repost to help others in your network
📩 Save for later

The DeepSeek Breakthrough That Just Ended the “Bigger Models” Era 

Honesty wins.

LLMs break because they bluff.
Flawless tone.
Faulty logic.

DeepSeek attacked the flaw.
Not the symptoms.
Not the scale.
The flaw.

✅ What DeepSeek built

A self-verifying pipeline
that rewrites how models think.

Verifier scans a proof
and flags broken logic.
Meta-Verifier checks the verifier
to prevent hallucinated judgment.
Generator writes a proof
and critiques itself step-by-step.

The model earns rewards
for admitting mistakes.
Not hiding them.
Not bluffing.

Honesty becomes the shortcut.
Not luck.

✅ Why it matters

Answer-only training hit a ceiling.
Correct answer isn’t reasoning.
And theorem proving has no numeric reward.

DeepSeek’s system lets models:
Audit their own logic.
Spot flawed steps.
Refine proofs over iterations.
Scale to harder problems
without any human labels.

Reasoning that improves itself.
Finally.

✅ The results

IMO 2025: solved 5 of 6.
CMO 2024: 4 full + 1 partial.
Putnam 2024: 118 of 120
(humans peak at 90).

It even outperforms
GPT-5-Thinking-High
and Gemini 2.5 Pro
across all major math tasks.

The first open model
with gold-level theorem proving
in natural language alone.

✅ The shift begins

Scaling isn’t the frontier.
Truthfulness is.
Models must audit themselves
like mathematicians.
Not magicians.

DeepSeek proved it’s possible.
And the field just changed.

➕ Follow Ghadeer for more insights
♻️ Repost to help others in your network
📩 Save for later

DeepSeekMath_V2

They compressed 4,000 words into a single image. 
The AI still remembered everything

This isn’t OCR it’s optical memory.

AI models choke on long documents.
Too many tokens. Too much cost.
DeepSeek just broke that limit.

✅ DeepSeek-OCR turns text into images and compresses it 10×.
It reads those “visual tokens” back with 97% accuracy.
Words became pixels. Meaning stayed intact.

✅ Even at 20× compression, it keeps 60% precision.
That’s like shrinking a whole book into one image 
and the AI still knows every page.

✅ Powered by a new DeepEncoder.
It fuses two attention systems,
reducing 4,096 image patches to just 256 tokens.

✅ It beats every major OCR model.
Fewer tokens than GOT-OCR2.0.
Faster than MinerU2.0 by 8×.

✅ Built for industrial scale.
1 GPU = 200,000+ pages/day.
20 nodes = 33 million pages/day.

✅ And it goes beyond text.
It reads charts, formulas, geometry 
even 100+ languages.

This isn’t about reading anymore.
It’s about seeing to remember.
DeepSeek-OCR just turned vision into the new language of memory.

Check out the repo: https://lnkd.in/dFAP8DTf

➕ Follow Ghadeer for more insights
♻️ Repost to help others in your network
📩 Save for later

I had a painful realization last week while reviewing a broken agent workflow…

Most engineers still think “seeing” is enough.
But visual understanding is the new baseline.

If your model can’t interpret charts, diagrams, objects, layouts, and logic the way you do…
you’re already behind.

2025 is the year multimodal literacy becomes a non-negotiable skill.

Here are the 10 VLMs every serious AI engineer must master or risk playing catch-up while others build the future:

1. Gemini 2.5 Pro (Google)
Understands text, image, audio, and video like a human brain.
A leap toward unified multimodal intelligence.

2. InternVL3-78B
Open-source powerhouse breaking records in multimodal reasoning.
Rapidly becoming the default for serious builders.

3. Ovis2-34B
Aligns text and visuals with lightning precision.
A top pick for tight visual-text alignment.

4. Qwen2.5-VL-72B
The backbone of the new wave of open multimodal AIs.
Trusted for stable, large-scale deployments.

5. o3 (OpenAI)
The new reasoning king with math, science, and visuals mastered.
Sets a fresh bar for structured problem-solving.

6. GPT-4.1
Reads charts, counts objects, and decodes visual logic flawlessly.
Perfect for production-grade visual workflows.

7. Claude Sonnet 4
Anthropic’s creative and analytical visual thinker.
Excels at nuance and safer interpretability.

8. Kimi-VL-A3B-Thinking-2506
Thinks faster and reasons smarter.
Cuts thinking time by roughly 20%.

9. Gemma-3-27B-it (Google)
Small model. Big performance.
Punches far above its weight class.

10. aLlama-3.2-90B-Vision-Instruct (Meta)
Sees, understands, and describes the world with precision.
Meta’s most capable vision-native model yet.

AI that sees is old news.
AI that understands is the real edge.
And that future has already arrived.
Don’t be the one catching up.

➕ Follow Ghadeer for more insights
♻️ Repost to help others in your network
📩 Save for later

I realized something uncomfortable this week about the way we talk about agents on LinkedIn…

I’ve been looking at a lot of “agent” posts here recently.

And the more I see, the more I notice the same pattern:

Most of what we call “agents” today aren’t agents at all.
They’re chatbots with longer prompts.

No planning.
No adaptive loops.
No tool orchestration.
No real memory.
No error handling.
No autonomy.

Just a nice wrapper and a big claim.

Then I came across Packt’s book, Building Agentic AI Systems, and honestly… it was refreshing.

It explained agents in a way that cuts through the noise completely.

The biggest thing that stuck with me:

Agents aren’t prompts.
Agents are systems.

Real agents look more like architectures than chat interfaces.

They need:

🔹 A reasoning loop that doesn’t fall apart on step 2
🔹 A planning module that adapts when the task changes
🔹 A memory stack (vector + episodic + long-term)
🔹 Tool orchestration that survives messy inputs
🔹 Error handling + verification
🔹 Human override channels
🔹 Clear autonomy boundaries
🔹 Optional multi-agent setups

This is the difference between:

❌ a fun demo that goes viral on LinkedIn
and
✅ a system that actually survives in a real deployment

Because when you zoom out, the truth is simple:

If you’re building agents, you’re really building architectures.
Not magic prompts.
Not “AI personalities.”
Not one-click workflows.

Systems.

It’s a shift I think more of us need to internalize myself included.

If you’re working in AI engineering, this book is worth a read.
Not because it’s trendy, but because it’s practical.

Link to book: https://lnkd.in/dPUvBWqr

What’s the biggest gap you’re seeing in today’s “agent” hype?
I’d love to read your thoughts.

♻️ Repost to help others in your network

The Weaviate Playbook Just Exposed the Biggest Lie in Prompt Engineering.

For the last two years, everyone has been obsessed with prompts.

Better wording.
Better templates.
Bigger context windows.
More “magic” instructions.

But the Weaviate Context Engineering Playbook makes one thing painfully clear:

Prompts were never the real bottleneck.
Context architecture was.

Most AI systems don’t break because the model is weak.
They break because the meaning you feed the model is unstructured, noisy, inconsistent, or irrelevant.

Here’s the truth the playbook nails:

🔹 Your chunking is wrong (too big = noise, too small = no meaning)
🔹 Your embeddings aren’t aligned to the task
🔹 Your retrieval is optimizing for similarity, not relevance
🔹 Your context has zero formatting contracts
🔹 Your system treats context like text. Not like an API

And the worst part?

Teams keep tuning prompts to hide these failures.
They’re treating symptoms while ignoring the disease.

The top 1% of AI teams are already shifting to this new paradigm:

🔹 Context Engineering > Prompt Engineering
🔹 Architecture > Tricks
🔹 Meaning > Tokens

Once you fix context, the hallucinations drop.
The model becomes predictable.
And your entire AI stack goes from “occasionally brilliant” to “consistently reliable.”

This isn’t just a playbook.
It’s a correction to the entire industry.

Prompt engineering didn’t scale.
Context engineering will.

➕ Follow Ghadeer for more insights
♻️ Repost to help others in your network
📩 Save for later

taplio

The Battle

Alex Razvant

Ghadeer A.

Followers

Posts

Engagements

Likes

Performance ScoreAlex Razvant's LinkedIn Performance

Performance ScoreGhadeer A.'s LinkedIn Performance

Ready to grow your LinkedIn presence?

taplio

The Battle

Alex Razvant

Ghadeer A.

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 ScoreGhadeer A.'s LinkedIn Performance

Ready to grow your LinkedIn presence?

Performance ScoreAlex Razvant's LinkedIn Performance