Rami Krispin's Data Science Channel

OpenWhispr: Use Your Voice For Everything This video from NeuralNine introduces OpenWhispr, a cross-platform voice-to-text app that supports private local transcription with Whisper or NVIDIA Parakeet, along with cloud and bring-your-own-key options. https://www.youtube.com/watch?v=PKzCW5OpMVM

Issue 104 is out! This week's agenda: 🔹 Open Source of the Week - the agent-top project by Kannan Kalidasan 🔹 New learning resources - Docker Sandboxes, local real-time speech, TrueForge, and pure Python AI agents 🔹 Book of the week - Designing AI Interfaces: Design Principles for Creative and Autonomous AI by Louise Macfadyen https://ramikrispin.substack.com/p/agent-top-designing-ai-interfaces

I feature an open-source project every week in my newsletter, and last week's pick is a really cool project: Archify. Archify is a Node.js rendering and validation system that turns a codebase or system description into an interactive system map. It uses a typed JSON intermediate representation and deterministic compilation to keep diagrams reproducible and make validation failures easier to repair. Key features include: ✅ Architecture, workflow, sequence, data-flow, and lifecycle diagrams ✅ Typed JSON diagram source ✅ Schema, layout, HTML, SVG, route, and label validation ✅ Grounded search and relationship tracing ✅ Validated snapshot comparison ✅ Interactive themes, focus controls, and guided stories ✅ Self-contained HTML, PNG, SVG, WebM, and share-card exports More details are available here: https://github.com/tt-a1i/archify License: MIT ♻️ Please share if you find it useful 📌 Subscribe to my newsletter: https://ramikrispin.substack.com/

BreezeTTS2 - Local Real-Time Voice This video from Sam Witteveen examines BreezeTTS2, a 3B open-weight text-to-speech model for local real-time inference. The 14-minute video covers voice design and direction, voice events, multilingual generation, latency, benchmarks, licensing, voice cloning, and streaming. https://www.youtube.com/watch?v=xDHD09fDUkQ

A new tutorial in my Docker 🐳 101 series — this time focusing on building Docker images 👇🏼 I recently started a series of Docker tutorials for AI/ML developers. The first sequence walks through the core Docker workflow: requirements → Dockerfile → docker build → docker run The 4th tutorial in the series focuses on the build stage and covers: ✅ How the Dockerfile, build context, and image fit together ✅ How to build and tag an image with docker build ✅ What happens when Docker resolves a base image through Docker Hub ✅ How to confirm that the finished image exists with docker images The next tutorial in the sequence will focus on working with a container registry. 🔗 https://theaiops.substack.com/p/how-to-build-a-docker-image-from?r=1x99er&utm_medium=ios

ssue 103 is out! This week's agenda: 🔹 Open Source of the Week - The Archify project 🔹 New learning resources - Docker debugging, RecBole recommender systems, agentic harnesses, visual data structures and algorithms, reinforcement learning for unverifiable tasks, and LLM inference system design 🔹 Book of the week - Vector Databases: A Practical Introduction by Nitin Borwankar https://ramikrispin.substack.com/p/the-archify-project-vector-databases

GLM 5.3 can run locally if you have an appropriate machine with ample RAM (e.g., a 256GB Mac or a system with sufficient RAM/VRAM). Here is a guide from Unsloth: https://unsloth.ai/docs/models/glm-5.3

The MLX fast is a new cool community project with the goal of finding the best setting for running different LLM locally on Apple silicon 👇🏼 https://www.yukon.org/mlxfast

The new VScode version comes with new markdown editor functionality. One of the nice features is editing the markdown directly in preview mode 👇🏼 https://youtu.be/7uyRMACA_pM?is=xQ1MI9QdWQdfv5H-

I feature a data science book every week in my newsletter, and this week's pick focuses on Imbalanced Data: Myths, Mistakes and Modern Solutions by Soledad Galli, PhD. The book challenges the common assumption that class imbalance must be fixed through resampling. It builds an evidence-driven approach to classification around model evaluation, decision thresholds, probability calibration, real-world error costs, and careful use of sampling techniques. Topics include: ✅ Class imbalance foundations and the factors that affect model performance ✅ Classification metrics and how to match them to the decision problem ✅ Decision thresholds that reflect practical requirements ✅ Probability calibration and recalibration in Python ✅ Cost-sensitive learning and the costs of different errors ✅ Undersampling, data cleaning, and the tradeoffs of removing observations ✅ Oversampling, SMOTE, and the bias synthetic sampling can introduce ✅ Using AI to build and audit an imbalanced-learning pipeline It is a practical resource for data scientists and machine learning engineers who want to make better decisions about metrics, thresholds, calibration, and resampling. More details are available here: https://www.trainindata.com/p/imbalanced-data-myths-mistakes-solutions-book ♻️ Please share if you find it useful 📌 Subscribe to my newsletter: https://ramikrispin.substack.com/

My weekly newsletter is out 👇🏼 https://ramikrispin.substack.com/p/skills-for-real-engineers-learning

The first step in the Docker 🐳 workflow, after defining the environment requirements, is to define the Dockerfile. In a nutshell, a Dockerfile is a text file containing the instructions Docker uses to build an image. It is one of the most important components of the Docker workflow. A well-designed Dockerfile can save development time, shorten image builds, and reduce infrastructure costs throughout the application lifecycle. This is the second tutorial in the Docker 101 series focusing on the Dockerfile. We will learn how the Dockerfile works by using a practical example: dockerizing a FastAPI application. This includes: ✅ Defining requirements ✅ Setting up the Python version ✅ Defining a virtual environment and installing the required libraries ✅ Copying the FastAPI script file to the image and making it executable The next tutorial in this sequence will focus on the next step in the workflow - the build. https://theaiops.substack.com/p/how-to-write-your-first-dockerfile

I feature an open source project every week in my newsletter, and this week's pick is Unsloth Studio. Unsloth Studio is a browser-based interface for running and training models locally without building the workflow in code. It brings model discovery, chat, dataset preparation, fine-tuning, monitoring, export, and serving into one interface for Mac, Windows, and Linux. Key features include: ✅ A local browser-based interface for running and training models ✅ Model discovery and chat with GGUF and Safetensors models ✅ Side-by-side model response comparison ✅ Tool calling, web search, and Python or Bash execution ✅ No-code fine-tuning configuration and training monitoring ✅ Data Recipes for turning files into training datasets ✅ GGUF, Safetensors, and LoRA model exports ✅ OpenAI-compatible API serving for applications and agents More details are available here: https://github.com/unslothai/unsloth License: Apache 2.0 📌 Subscribe to my newsletter: https://ramikrispin.substack.com/

Issue 101 is out! This week's agenda: 🔹 Open Source of the Week - Unsloth Studio 🔹 New learning resources - Hermes desktop workflows, ML system design, and Docker Sandboxes 🔹 Book of the week - Imbalanced Data: Myths, Mistakes and Modern Solutions by Soledad Galli https://ramikrispin.substack.com/p/unsloth-studio-imbalanced-data-issue

Getting Started with Docker Sandboxes 🚀 This tutorial from Docker walks through setting up an isolated development environment for AI-assisted coding. The five-minute tutorial covers: ✅ Installing the Sandbox CLI ✅ Authenticating with Docker ✅ Launching a sandbox ✅ Mounting a workspace ✅ Using the same sandbox with Claude and Codex 📽️: https://www.youtube.com/watch?v=9jfk1fdhXZo

ML System Design for Data Scientists 🚀 This video from Andres Vourakis, a Senior Data Scientist at Nextory, walks through a production ML system design framework. The 17-minute video covers: ✅ Six-step design framework ✅ Four production requirements ✅ Trial-to-paid conversion example ✅ Layer-by-layer system design ✅ Interactive ML system map 📽️: https://www.youtube.com/watch?v=t3W3TJrNUnQ

The first tutorial in this series is out and focuses on the motivation for using Docker. https://theaiops.substack.com/p/why-ai-developers-should-learn-docker

Docker 101 🐳👇🏼 Want to learn Docker? I am starting a new series of tutorials focused on Docker for ML/AI developers. The first tutorial is out. Docker isn't as hard as it sounds, but it has a steep learning curve, and the goal of this series is to smooth it out. What are the benefits of Docker? Reproducibility. Reproducibility is critical in any form of software development, mainly when you ship your code to another machine (e.g., deploy or have teammates run it). This Docker series is going to cover: ✅ What is Docker and when should you use it ✅ Containerizing a minimal Python AI application. ✅ Writing Dockerfiles that remain fast and maintainable as a project changes. ✅ Building reproducible environments with Docker Compose and Dev Containers. ✅ Running local models with Docker Model Runner. ✅ Designing container boundaries for agents, MCP tools, and Docker Sandboxes. ✅ Testing, securing, and publishing multi-platform images through CI.

Meta Muse Glimmer 🚀 Meta released a 30B open-weight agentic model under Apache 2.0 for local workflows on consumer hardware. It supports tool use, long-horizon reasoning, failure recovery, text+image input, adjustable reasoning effort, and 100+ languages. A roughly 4-bit version fits under 20 GB. More details: https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model