Data Science Dojo

The May 5th cohort is 6 days away. 10 weeks. 12,000+ alumni. Spots are filling fast. The Agentic AI Bootcamp is back and it covers the full stack of what matters right now. Week 1, you're wrapping your head around how Transformers actually work. Week 3, you're building with LangChain and setting up your first vector database for a RAG pipeline. By week 5, you're not just prompting. You're engineering context and designing systems using Agentic design patterns used in production. Week 7 hits and you're implementing Model Context Protocol and learning how agents actually talk to each other. Week 10? You ship a multi-agent LLM application. From scratch. In 10 weeks. That's what the Agentic AI Bootcamp cohort starting May 5th looks like. 3 hours a week, structured to take you from foundations to deployment. Register now before it's too late: hubs.la/Q04dShn50 #agenticai #rag #aibootcamp #vectordatabases

💡 Better AI outputs don't start with better prompts — they start with better context. The shift from "prompt engineering" to context engineering is one of the most important mental model upgrades you can make when building with LLMs. Here's exactly how a well-architected context pipeline works: 🔹 Step 1: Gather your inputs. Every request draws from four sources: the user prompt (what they asked), the system prompt (standing instructions that shape behavior), memory (past prompts and responses to maintain continuity), and a vector database (semantically relevant documents retrieved at query time). These aren't optional extras — they're the raw material the model thinks with. 🔹 Step 2: Augment with web search. For anything time-sensitive or outside the model's training data, a web search runs in parallel, pulling results relevant to the user's query before the model ever sees it. 🔹 Step 3: Merge everything into one context. All those sources collapse into a single unified context window. The model doesn't know which source is which — it just sees one coherent block of information to reason over. 🔹 Step 4: Generate an intermediate output. The LLM produces a first-pass response. This is deliberately not the final answer. 🔹 Step 5: Critique and revise. A Reviewer/Critic model (or the same model with a separate critique prompt) evaluates that output. If revision is needed, it loops back. If not, the final output ships. The loop at step 5 is what separates fragile pipelines from robust ones. One-shot generation works until it doesn't. Build in the critic. What part of this pipeline is hardest to get right in your experience? #ContextEngineering #LLM #AgenticAI #RAG #PromptEngineering #AIEngineering #LLMOps #GenerativeAI

The July 14th cohort is coming up soon. 10 weeks. 12,000+ alumni. Spots are filling fast. The Agentic AI Bootcamp is back and it covers the full stack of what matters right now. Week 1, you're wrapping your head around how Transformers actually work. Week 3, you're building with LangChain and setting up your first vector database for a RAG pipeline. By week 5, you're not just prompting. You're engineering context and designing systems using Agentic design patterns used in production. Week 7 hits and you're implementing Model Context Protocol and learning how agents actually talk to each other. Week 10? You ship a multi-agent LLM application. From scratch. In 10 weeks. That's what the Agentic AI Bootcamp cohort starting July 14th looks like. 3 hours a week, structured to take you from foundations to deployment. Register now before it's too late: hubs.la/Q04mh6kk0 #agenticai #rag #aibootcamp #vectordatabases

Read more here: sakana.ai/fugu-release

🚨 Sakana AI just launched Fugu — a system that matches Fable's performance without even being a frontier model. The idea is simple on the surface: instead of one giant model handling everything, Fugu coordinates a pool of existing models and routes work between them. But the interesting part is how it decides who does what. 🔹 Internally, every request gets split across three roles: a Thinker (plans the approach), a Worker (executes it), and a Verifier (checks the output) 🔹 A lightweight coordinator model — trained using an evolutionary algorithm, not hand-coded rules — assigns those roles dynamically based on the task 🔹 A second system called Conductor, trained with reinforcement learning, figures out how agents should communicate and pass context to each other 🔹 The whole thing is recursive: Fugu can call instances of itself as sub-agents for complex tasks From the outside, you call one OpenAI-compatible API endpoint. The orchestration is invisible. Two variants shipped: standard Fugu for everyday tasks and Fugu Ultra for multi-step, high-accuracy workloads. Fugu Ultra matches Claude Opus 4.8 and GPT-5.5 on several benchmarks — without either in its agent pool, since both are subject to export controls. That's the architecture's main claim: collective intelligence that doesn't depend on any single provider. #SakanaAI #AIAgents #MultiAgentSystems #LLM #AIEngineering #AgenticAI

Meet our speaker for this week's live session: Dan Ndombe 👋 Dan is a Staff Developer Success Advocate at Docker, where he helps developers build and ship software faster. He's a two-time founder and an engineer-turned-product manager, with prior experience at Netflix, Pinterest, and Calm. Join Dan on June 24 at 1:00 PM PT for "Running LLM Agents Safely: Hands-On with Docker Sandboxes" - a hands-on session on isolating and securing autonomous AI agents using Docker Sandboxes. 👉 Register now: hubs.la/Q04mf2NK0 #LLMAgents #DockerSandbox #AISecurity #Docker #AIAgents #DataScience

Link to the repo: github.com/multica-ai/and…

📢 A single text file inspired by Andrej Karpathy picked up 175k stars on GitHub. That's the whole story. The file is a CLAUDE. md, a set of instructions you drop into your project to make AI coding assistants behave better. It's based on Karpathy's public complaints about how these tools actually fail in practice: ❌ They make assumptions and keep going instead of asking ❌ They write 1,000 lines when 100 would do ❌ They touch code they were never asked to touch ❌ They "fix" things that weren't broken The file gives the AI four rules to follow: 🔹 think before writing any code, 🔹 keep solutions as simple as possible, 🔹 only change what was actually asked, 🔹 and verify the result against a clear success condition. It works with Claude Code, Cursor, and most other AI coding tools. One curl command installs it. #ClaudeCode #AIEngineering #DeveloperTools #LLM #OpenSourceAI

One URL change turns a research paper into running code. Swap arxiv for autoarxiv and an AI agent from alphaXiv takes it from there. Here's what it actually does: 🔹 Reads the abstract, the claims, and any linked GitHub repo 🔹 Clones the codebase and handles the usual setup problems: broken dependencies, missing paths, environment config, hardware assumptions 🔹 Designs a scaled-down version of the experiment — smaller model, fewer steps, one GPU instead of a cluster — just enough to test whether the headline result holds 🔹 Runs it live with full logging, so loss curves, metrics, and training progress are all visible as it happens What comes back is a clear answer on whether the minimal run matches what the paper reported, plus an estimate of what a full replication would cost in compute and time. This moves reproduction from a weekend of debugging to a single URL change. Video credits: @askalphaxiv/x #AIResearch #MachineLearning #ReproducibleResearch #AIEngineering #OpenScience

Link to the repo: github.com/bytedance/deer…

🚨 The company behind TikTok, just open-sourced DeerFlow 2.0 — an AI agent framework that hit #1 on GitHub Trending. Here's what it does: 🔹 Breaks big tasks into smaller pieces and hands them to multiple AI agents working at the same time 🔹 Remembers what you've worked on before, even across separate sessions 🔹 Runs code in a safe, isolated environment instead of directly on your machine 🔹 Only loads the tools and instructions it actually needs for a given task, so it doesn't get bogged down The whole thing is free and open source under the MIT license. If you're experimenting with AI agents that need to handle multi-step work on their own, this one's worth a look. #DeerFlow #AIAgents #OpenSourceAI #AIEngineering #ByteDance

Explore our curriculum here: hubs.la/Q04m4PHJ0 For a detailed guide on our bootcamps, join our free info session: linkedin.com/events/7473819…

❓ Why do multi-agent systems exist? Context window limits in earlier models made single-agent setups fall apart on complex, multi-step tasks. That's the starting point Zaid Ahmed used in Session 7 of our Agentic AI Bootcamp. Here's what the session covered: - Reflection, tool use, and planning patterns, recapped before moving into multi-agent collaboration - MCP, A2A, and ACP protocols, and how they cut the manual work of extracting data from tool calls - A hierarchical multi-agent system with a research agent and a chart generator agent, wired together with LangChain Expression Language - Deep agents that spawn their own sub-agents and share artifacts through a virtual file system, sandboxed in isolated environments like Azure Containers - Prompt caching and semantic caching, with Redis handling the query-answer pairs Participants also built a prompt generator from scratch using LangGraph, the kind of hands-on exercise that turns "reflection pattern" from a slide into something you've actually shipped. One takeaway worth repeating: as agents get more autonomy, Human-in-the-Loop checkpoints stop being optional. Our next Agentic AI Bootcamp starts July 14, 2026. 10 weeks, instructor-led. Link in the comments. #AgenticAI #MultiAgentSystems #LangGraph #AIEngineering #MCP

A new paper just pulled apart Claude Code's source code, and one number stood out: only 1.6% of the codebase is actual AI decision logic. The other 98.4% is just plumbing. Researchers at MBZUAI looked at the actual code behind Claude Code and compared it to OpenClaw, another open-source AI assistant, to understand how these tools get built in practice. Here's what they found: 🔹The core loop is almost embarrassingly simple - the model picks an action, the tool runs it, repeat. Everything else exists to support that one loop. 🔹 People approve 93% of permission prompts without really checking them, so just asking "are you sure?" doesn't actually keep anyone safe 🔹 The system trims and compresses the conversation five different ways before every single response, just to avoid running out of room to think 🔹 When one task spins off a sub-task, only the summary comes back - not the whole conversation, so things don't get cluttered The idea behind all this is simple: assume the AI will sometimes get things wrong or get tricked, and stop trying to make it smarter. Build strong rules and checks around it instead. The AI suggests what to do. The surrounding system decides whether to actually let it happen. That's actually the opposite of where a lot of AI tools are headed. Many teams are pouring effort into making the AI plan better. This paper suggests the bigger win is building a solid system around it. If you're building anything with AI agents, this is worth sitting with: the boring infrastructure work might matter more than the clever AI part. #AI #AIAgents #ClaudeCode #AIResearch

🤖 RAG gives LLMs access to your data. Agentic RAG gives them judgment about what to do with it. The difference shows up in how each system handles uncertainty. Standard retrieval-augmented generation runs a fixed pipeline: encode the query, search the vector database, retrieve similar documents, generate a response. It works well when the query is well-formed and the right context already exists in one place. Agentic RAG adds a layer of reasoning at almost every step — rewriting the query before retrieval, deciding whether the initial results are actually sufficient, choosing between multiple source types (vector DB, APIs, live web), and evaluating whether the final answer is relevant before surfacing it. If any of those checks fail, it loops back rather than proceeding. That feedback loop is what makes it genuinely different, not just incrementally better. A standard RAG pipeline doesn't know when it's about to give a bad answer. An agentic one can at least ask the question. #agenticai #agenticaibootcamp #aibootcamps #rag #agenticrag #aiengineering

Explore our curriculum here: hubs.la/Q04m4PbX0 For a more detailed guide on our bootcamps, register now for our free information session: linkedin.com/events/7473819…

📢 Our LLM Bootcamp cohort fine-tuned a model and shipped a deployed agent in the same week. Here's what else went into the 5 days: 🔹 A fine-tuned LLaMA2-7B (4-bit quantized, using LoRA) — taught by Kartik Talamadupula, Distinguished Architect at Oracle 🔹 A full RAG pipeline in LangChain, from document loaders to LCEL 🔹Hybrid and generative search pipelines on vector databases 🔹 Defenses against prompt injection and data leaks in multi-agent systems — taught by Leonid Yankulin, Senior Developer Relations Engineer at Google 🔹 A deployed capstone, ranging from RAG chatbots to MCP and deep agents — closed out by Dan Ndombe, Staff Developer Success Advocate at Docker, on running LLM agents safely Raja Iqbal opened the week with where the LLM landscape actually stands, and Luis Serrano, Founder of Serrano Academy, broke down transformer architecture from the ground up. From theory to practical, we have it all. Our 40-hour LLM Bootcamp runs in-person and online. Link in the comments. #LLMBootcamp #RAGPipeline #AIAgents #FineTuning #AIEngineering

An agent that can act on its own is powerful. It's also where things can go wrong. 🔒 Docker Sandboxes give your AI agents an isolated environment - scoped file access, restricted network calls, capped resources, and a fresh start every run. Join Dan Ndombe (Docker) live on June 24 at 1:00 PM PT to build a secure agent runtime from scratch. 👉 Register now: hubs.la/Q04m41v70 #LLMAgents #DockerSandbox #AISecurity #Docker #AIAgents #MachineLearning #BuildWithAI #MLOps

🎙️What are humans doing with the time they are reclaiming using AI? Mark Cavage, COO & President Docker, doesn't subscribe to the "AI ends work" narrative. Like every technological revolution before it, AI should unlock a new level of productivity — but only if we actually use the time we reclaim wisely. The transition will be hard. But once the dust settles, we should come out collectively better off. 🔹 Why job panic misses the actual pattern of how revolutions work 🔹 How the time AI gives back gets swallowed by burnout instead of progress 🔹 What a real productivity gain looks like when the transition ends Full episode streaming now on YouTube, Spotify and Apple Podcasts. Stream on your favourite platform: hubs.la/Q04lqkcL0 #AgenticAI #AIProductivity #FutureOfWork #Automation #FutureofDataAndAI

🚨 Codex can now learn a workflow by watching you do it once, then turn it into a reusable skill. Record and Replay is live in the Codex app's Plugins menu. It captures your mouse clicks, typed text, and on-screen content while you complete a task, then drafts a skill from what it observed. What the generated skill includes: 🔹 When to use the workflow 🔹 What inputs it needs 🔹 The steps to follow 🔹 How to verify the result Once it's saved, Codex can run that skill again with Computer Use, browser actions, or connected plugins, swapping in new inputs like a different file or date range. It's built for the workflows that are faster to show than to describe in a prompt: filing an expense, creating a correctly formatted issue, or pulling a recurring report. If you're maintaining agent workflows, this is worth testing on anything repetitive where the steps are stable. It skips the part of prompt writing where you try to put muscle memory into words. #CodexAI #ComputerUse #AIAgents #DeveloperTools

📢 Everyone's been throwing around "agent loops" lately, but if you're still not sure what one actually looks like in practice, Loop Library just made that easier for you. Matthew Berman created Loop Library, a site for reusable agent prompts for engineering, research, evaluation, and ops work, and it's open to contributions from anyone using loops in production. Each loop follows the same structure: - A trigger that starts it - An action the agent takes - A proof step to verify the work - A stopping condition so it knows when to quit That stopping condition is the part most prompt collections skip, which is why agents often keep "fixing" things long after the real work is done. Loops like the 100% test coverage loop and the production error sweep both report back automatically once the verify step passes, instead of running indefinitely. If you're running agent loops yourself, the verify/stop format here is worth stealing even before you touch the specific prompts. #AgentLoops #AIEngineering #PromptEngineering #AgenticAI #DevTools