GitHub radar

Top 5 GitHub Repos This Week

This week I noticed a clear split: one half of the trending list was all about shrinking AI to the edge — literally running it on an $8 chip — while the other half was about rethinking the TypeScript toolchain from scratch. Five repos worth opening, zero filler.

01slvDev/esp32-ai 1.5kPython

I ran into this while browsing and immediately ordered an ESP32-S3 — the idea of a 28.9M-parameter language model running on an $8 chip at 9.5 tokens per second felt almost absurd. Pulled the repo, checked the code — it's clean, real implementation, not a demo wrapper. The trick is 4-bit quantization with Per-Layer Embeddings borrowed from Google's Gemma, storing 25M of those parameters straight in flash with only 512KB SRAM on the device. It generates short coherent stories locally, no server, no Wi-Fi required, and what I liked most is the author explains the entire architecture with actual numbers in the README — no hype, just engineering.

Why a vibe-coder should care

If you're building edge AI products — IoT devices, offline assistants, hardware demos — this is the clearest proof-of-concept I've seen that real inference at the edge is no longer theoretical.

Open on GitHub
02Jakubantalik/thinking-orbs 1.1kTypeScript

I installed this locally in literally two minutes — npm install thinking-orbs — and six animation states were ready to use: Working, Searching, Solving, Listening, Composing, Shaping. Each one is actually thought through, not just a spinner variant. What struck me most is it's built on pure 2D canvas with no ctx.filter, no SVG filters, no WebGL — the author clearly uses this themselves and cared about performance. Respects prefers-reduced-motion, auto-pauses when the tab is hidden, ARIA labels out of the box — exactly the small things that separate a quality library from a weekend project.

Why a vibe-coder should care

If you're building an AI agent or chat interface and your loading state is still a generic spinner, this drops in instantly and makes the UI feel like a real product.

Open on GitHub
03mshumer/Claude-of-Duty 746JavaScript

I'll be honest — I opened this expecting another vibe-coding toy. But the code is 55,000 lines across 11 real subsystems: HDR pipeline with cascaded shadow maps in a sampler2DArray, a custom physics engine using binned-SAH BVH, AI soldiers with navmesh pathing, Web Audio synthesis without a single prerecorded sound file. What I found most interesting in the README is mshumer's observation that parallel agent fanout actually performed worse than sequential single-owner passes — that's a non-obvious, honest insight about agentic development. FPS improved from 12-17 to 28-30 across three refinement rounds.

Why a vibe-coder should care

Beyond the game itself, this is one of the most honest writeups of what multi-agent AI development actually looks like in practice — including what failed and why sequential beats parallel.

Open on GitHub

I stumbled across this and immediately started counting how much manual work it replaces. You feed it Chinese story text or a sequence of hand-drawn images and it outputs a 3:4 vertical video with handwritten captions, black-and-white plate reveals that transition to color illustrations, and optional page-flip effects — all driven by Remotion with a TypeScript and FFmpeg pipeline underneath. What surprised me is that it's packaged as an agent skill for Claude Code, Codex, and Kimi Code, so you can describe a scene in natural language and get an animation back. The output is a silent H.264 MP4 at 1080×1440, which is a thoughtful choice — leaves room for voiceover in post.

Why a vibe-coder should care

For anyone building AI content pipelines or animated explainers, this shows concretely how an agent skill can turn text into real video output with minimal glue code.

Open on GitHub
05vercel-labs/scriptc 633TypeScript

From the first look it's clear this isn't another toy 'compile TS to exe' project — Vercel Labs is taking this seriously. scriptc compiles TypeScript to a native binary via LLVM or C, with no Node.js or V8 bundled inside. Startup time is 2.4ms versus Node's 47ms, binary size 170-200KB, memory usage 1-4MB versus Node's 67-116MB. It already supports async/await via stackful fibers, monomorphized generics, JSON with runtime-validated casting, fetch, and http/https — and uses differential testing to verify byte-for-byte parity with Node.js. This is beyond the proof-of-concept stage.

Why a vibe-coder should care

For CLI tools, serverless functions, or edge deployments where Node startup latency and memory overhead are real bottlenecks, this is worth serious attention — especially with Vercel Labs behind it.

Open on GitHub