Rami Krispin's Data Science Channel

Tool, skill, or subagent? This great workshop from Anthropic focuses on the different tooling available to build agents for Claude code. The main components of an agent are: ✅ Tool — a deterministic function the agent calls when it needs a specific action ✅ Skill — a markdown file the agent reads on demand to load a procedure or domain knowledge ✅ Subagent — a separate Claude Code agent with its own prompt and tools, invoked from the main one This workshop focuses on general agent architecture, best practices for building agents using the available functionality, setting evals, and maintaining agents as the scope grows. 📽️: https://www.youtube.com/watch?v=mWvtOHlZM-I

freeCodeCamp dropped a new full course: Build an AI Agent in Telegram with Vercel, Cursor, Composio. Start from the Vercel AI SDK chatbot template, connect Gmail via Composio, add long-term memory with Supermemory, ship through Telegram, and schedule jobs with Vercel Cron. Covers tool calling, context engineering, OAuth, memory, and deployment. 📽️: https://www.youtube.com/watch?v=hIh2O9OL69o

If you are building HTML slides, here is a great skill with great templates: https://github.com/zarazhangrui/beautiful-html-templates

How to use GLM 5.2 for free? According to the below tutorial, one of the free models in OpenCode are GLM 👇🏼 https://www.youtube.com/watch?v=jzjhRgya2gM

Here is how you can run GLM 5.2 locally, if you simply have a machine with 256GB to 810GB RAM 😅 https://unsloth.ai/docs/models/glm-5.2

What if every data source looked like a local folder to your AI agent? Mirage is a new open-source project from Strukto that lets you mount services like S3, Google Drive, Slack, Gmail, Redis, and many others into a single virtual filesystem. Instead of writing custom SDK integrations or MCP tools for every service, agents can use familiar filesystem commands such as cp, cat, grep, find, and wc across all connected resources. Why this is interesting: ✅ Around 50 built-in backends, including S3, GCS, Redis, Gmail, Google Drive, GitHub, Notion, Slack, PostgreSQL, MongoDB, Discord, Telegram, and more ✅ Bash-native pipelines that work across services (for example, copying files from S3 to local storage or searching Slack messages) ✅ Smart file handling based on file type (e.g., reading a Parquet file as JSON or extracting text from PDFs) ✅ Portable workspaces that can be cloned, snapshotted, and moved between machines ✅ Embeddable in Python and TypeScript applications without requiring a separate process ✅ Built-in caching with optional Redis support for sharing state across workers Mirage is licensed under Apache 2.0 and integrates with frameworks such as OpenAI Agents SDK, Vercel AI SDK, LangChain, Pydantic AI, CAMEL, OpenHands, and CLI-based coding agents like Claude Code and Codex. It currently supports macOS and Linux (using FUSE-based mounts). GitHub: https://github.com/strukto-ai/mirage Have you experimented with the filesystem abstraction for AI agents? I'd be interested to hear how it compares with MCP-based approaches. #AI #OpenSource #LLM #AIAgents

Fine Tune Gemme 4 Model Locally 👇🏼 This short tutorial walk through the fine-tuning process of Gemma 4 locally using Unsloth and Hugging Face using LoRA fully locally. No GPU or cloud needed. https://www.youtube.com/watch?v=CpgyqJcli10

If you're building with AI coding agents, it's time to learn Git worktrees. The biggest advantage of AI agents isn't that they write code—it's that they can work in parallel. The challenge? Multiple agents shouldn't be modifying the same working directory. That's where git worktree shines. Instead of switching branches or cloning the same repository multiple times, you can create separate working directories that all share the same Git history. Each AI agent gets its own isolated workspace to: ✅ Implement a feature ✅ Fix a bug ✅ Update documentation ✅ Experiment with a refactor ...all at the same time, without stepping on each other's changes. The result is a workflow that feels much more like collaborating with a team of developers than working with a single coding assistant. If you haven't used Git worktrees before, I highly recommend this excellent tutorial series by Net Ninja: https://www.youtube.com/watch?v=Vf_0QpLsFRs&list=PL4cUxeGkcC9iUtQh7Aja3TGfbdd7Z-K0W&index=1

improve is an open-source Agent Skill that audits your codebase and generates implementation plans for AI agents. The idea is simple: use a powerful model to analyze the codebase and create a detailed plan, then let cheaper models execute it. The workflow: 🔍 Recon — learns the project structure, coding conventions, and existing documentation. 🔎 Audit — analyzes the codebase across multiple dimensions, including correctness, security, performance, testing, technical debt, dependencies, documentation, and developer experience. ✅ Vet — validates findings to reduce false positives. 📊 Prioritize — ranks findings based on impact, effort, and confidence. 📝 Plan — generates a standalone implementation plan for each finding. A few features I found interesting: • Every finding links to the relevant file and line. • Plans include verification steps, expected outputs, and clear boundaries. • Plans are versioned against a specific Git commit to detect drift. • Can execute plans in isolated Git worktrees and review the results. • Supports focused audits (security, performance, tests, bugs), quick scans, and deep analysis. • Can export implementation plans as GitHub issues. • The skill never modifies your source code during the audit. Install: npx skills add shadcn/improve Repo: https://github.com/shadcn/improve

Build a local RAG pipeline with Ollama 🚀 New crash course from Python Simplified — turn any LLM into an expert with fast retrieval over your PDFs. Covers setup with LangChain + FAISS, chunking, embeddings, retrieval, and the chat model. https://www.youtube.com/watch?v=oZYlrooPgvs

Build 3 production AI agents in Python 🚀 New full course from Tech With Tim using Orkes's Agentspan framework — agents you can run in production, not toy demos. Covers the architecture, memory + tools, RAG, guardrails, and human-in-the-loop. https://www.youtube.com/watch?v=zFw19qGAeGo

While building my local LLM server with MLX, I came across the oMLX project, which already provides this functionality. Looks super awesome! https://github.com/jundot/omlx

7 agentic AI loops you can lift straight into your workflow 🔁 Matthew Berman's new ~10-minute walkthrough of his "Loop Library" — covering, among others, a sub-50ms page-load loop, an overnight doc sweep, an architecture-satisfaction check, and a logging-coverage loop. Library: https://signals.forwardfuture.ai/loop-library/ https://www.youtube.com/watch?v=F4a8aMLb678

One feature I found particularly useful is the CLI support. You can start, stop, or cancel transcription from scripts, making it easy to integrate into your own workflows. According to the project, Parakeet V3 can run at around 5× real-time speed on mid-range hardware, making it fast enough for everyday dictation. Install: • macOS: brew install --cask handy • Windows: winget install cjpais.Handy • Or download it from the project's releases page. Repo: https://github.com/cjpais/handy

Handy is a nice open-source tool for offline speech-to-text. Press a keyboard shortcut, start talking, and your transcript is pasted directly into whatever text field you're using. No cloud, no account, and no telemetry. What I like about it: ✅ Runs locally with Whisper or Parakeet ✅ GPU acceleration for Whisper, or CPU-optimized Parakeet V3 ✅ Uses Silero VAD to automatically stop when you stop speaking ✅ Global hotkey or push-to-talk ✅ Works on macOS (Intel + Apple Silicon), Windows, and Linux ✅ Supports custom Whisper GGML models ✅ Built with Tauri, Rust, and React ✅ MIT licensed

My newsletter is out! This week's agenda: 🔹 Open Source of the Week - DiffusionGemma 🔹 New learning resources - Gemma 4 12B MTP local test, Column-Level Data Lineage Engine From Scratch, DuoBench planner/implementer LLM pair benchmark 🔹 Book of the week - Large Language Models: The Hard Parts: Open Source AI Solutions for Common Pitfalls by Thársis Souza, PhD and Jonathan Regenstein https://ramikrispin.substack.com/p/diffusiongemma-column-level-data

ponytail — a skill that gets AI coding agents to write less code, not more 🚀 New open-source project from Dietrich Gebert. ponytail embeds a veteran senior developer persona into AI coding agents — "long ponytail, oval glasses, has been at the company longer than the version control" — so the agent stops over-engineering. The README's example: ask for a date picker, and without ponytail the agent installs flatpickr, writes a wrapper component, adds a stylesheet, and starts a discussion about timezones. With ponytail: <input type="date">. Key features: ✅ Decision ladder the agent follows before writing code — YAGNI first, then stdlib, then native platform, then installed dep, then one-liner, then the minimum that works ✅ Safety guardrails preserved — validation, error handling, security, and accessibility are explicitly never trimmed ✅ Intensity modes — lite, full (default), ultra, off ✅ Slash commands — /ponytail-review (current diff), /ponytail-audit (whole repo), /ponytail-debt (track deferred shortcuts), /ponytail-gain (impact scoreboard) ✅ Works with Claude Code, Codex, GitHub Copilot CLI, Gemini CLI, Cursor, Windsurf, Cline, Aider, Zed, and more ✅ MIT licensed Reported benchmark on a real Claude Code agent editing a FastAPI + React repo (Haiku 4.5, 12 tasks, n=4) vs. the same agent without the skill: ~54% less code on average (peak 94% on the date picker task), ~20% cheaper, ~27% faster, 100% safe. Get started in Claude Code: /plugin marketplace add DietrichGebert/ponytail /plugin install ponytail@ponytail Repo: https://github.com/DietrichGebert/ponytail

Self-healing CI/CD pipeline with AI 🛠️ New freeCodeCamp tutorial walks through wiring n8n, OpenAI, and GitHub Actions together so pipeline failures are automatically detected, analyzed, and resolved. Hands-on: a Node.js / Express demo app, a smoke test script, the GitHub Actions pipeline, and the n8n automation side. https://www.youtube.com/watch?v=vj68el9hRvU

Hermes agent architecture walkthrough 🛠️ New tutorial from Alejandro AO breaks down Hermes, his always-on AI agent — the agent loop, context construction and compression, gateway integrations with Telegram and Slack, SQLite-backed memory, and scheduled cron jobs. Includes a look at the context compression prompt that keeps long conversations workable. https://www.youtube.com/watch?v=n32qq7Kwzh0