Mike Hutu

2/2 No more hand‑crafted glue code—just import the library, point it at your graph, and you get a multimodal reasoning loop that can see, read, and query structured data in one pass. 🚀 github.com/NPC-Worldwide

🧵 1/2 The NPC‑Worldwide team just dropped npcpy, a Python toolkit that stitches together vision models, large language models, and knowledge‑graph back‑ends into a single, ready‑to‑run agent.

2/2 No more hopping between scattered notebooks and blog posts – everything you need lives in one place, from data wrangling scripts to model‑training notebooks. github.com/Yorko/mlcourse

🧵 1/2 If you’ve ever felt stuck between the math of ML and actually winning a Kaggle competition, this repo is the bridge you’ve been waiting for. The open‑source course bundles theory, code, and a full competition pipeline into a single, reproducible workflow.

2/2 As we move away from simple chat interfaces and toward autonomous agents that can execute code and access APIs, the surface area for these attacks is expanding at an alarming rate. github.com/CyberAlbSecOP

🧵 1/2 Building LLM applications without testing for prompt injection is essentially leaving the front door unlocked for your users.

@ivanfioravanti @Prince_Canuma @huggingface 2/2 The stack stitches together mlx‑vlm for vision, mlx‑audio for speech, and the tiny Gemma‑4 LLM for reasoning. The result is a sub‑100 ms round‑trip latency that feels native, not a remote‑server lag. github.com/Pocket-science

🧵 1/2 A few weeks ago @ivanfioravanti, @Prince_Canuma and @huggingface dropped a repo that finally lets you run a Reachy Mini with vision, TTS and STT entirely on a Mac. No cloud, no external GPU – just the Apple Silicon you already have.

3/3 github.com/ericosiu/ai-ma…

2/3 The ai-marketing-skills repo (ericosiu) bundles Claude Code “skills” that can run growth experiments, manage a sales pipeline, and even handle finance ops—all from Python. It’s not a single‑model demo; it’s a multi‑provider agent framework ready for production.

🧵 1/3 If you’re still clicking through spreadsheets to run SEO tests or chase leads, you’re missing the new wave of agentic workflows.

2/2 No more blob‑like outputs—each generated object comes with separate, functional parts that can be edited, rigged, or swapped out. github.com/wgsxm/PartCraf…

🧵 1/2 The single‑mesh era is over. PartCrafter, a NeurIPS 2025 paper now on GitHub, uses compositional latent diffusion transformers to break a 2D photo into a full‑blown, multi‑part 3D assembly.

3/3 github.com/plurai-ai/inte…

2/3 The new intellagent framework from plurai‑ai flips the script: it generates thousands of realistic synthetic interactions and runs them automatically against your agent, surfacing failures you’d never see in a handful of hand‑crafted prompts.

🧵 1/3 Manual QA can’t keep up with the combinatorial explosion of edge cases that break production agents.

2/2 No more stitching together separate scripts for research, UI mockups, or hand‑off – everything lives in one command line workflow. github.com/Owl-Listener/d…

🧵 1/2 The new "designer‑skills" repo just hit 1,008 stars on GitHub, and it’s not just a pretty demo. Owl‑Listener ships a full‑blown `op` CLI that lets you run a multi‑provider design agent from your terminal.

2/2 It’s Apache‑2.0, so you can ship it in commercial products without legal headaches. github.com/0xsline/StoryG…

🧵 1/2 Developers, stop spending hours hand‑coding ffmpeg pipelines. StoryGen‑Atelier lets you drop a text prompt and get a full storyboard plus a stitched video clip, all from JavaScript. The repo already has 755 stars and 121 forks, so the community is testing it hard.