IdeaTaste - A benchmark for research taste: let LLM agents propose research ideas, then check how many of them the research area actually went on to do.

Qwen3.8-27B EXL3 + DFlash2 speculative decoding on an AMD RX 7900 XTX (exllamav3-rocm): recipe, patches, benchmarks

HIP kernels for AMD Strix Halo (gfx1151) and a vLLM plugin that runs them on stock upstream vLLM: up to 31x faster prefill and ~2x decode at long context.

飲食熱量日誌 App。核心思想：讓 LLM 負責辨識品名與份量自動拆項後輸出結構化 JSON，並在後端接上可以成長的資料庫；熱量一律來自可核對的資料並由程式計算，每個數字都查得到出處。設計決策參考食物辨識研究及相關論文，以 Claude Code（Opus 5.5）協作開發。原為個人使用，以不含個人資料的全新 git 歷史開源發布。開發+實際使用後改進約三周，各項決策紀錄見docs/decisions/

TYSA (Teach Yourself Anything): an learning harness that finds where your understanding ends and teaches from there. Electron app built on Meta's Muse models.

Bulk classification and decision inference on one GPU: the content is read once, every question about it is scored in the same pass. OpenAI-compatible, Apache 2.0

闲亭 DeskPark — 简洁优雅的网页桌游平台：斗地主、掼蛋、德州扑克、四川麻将；单人人机与多人联机、房间聊天、像素虚拟形象、AI 托管人机，Docker 一键部署

An experiment-driven LLM engineering lab with 9 reproducible Python experiments exploring token economics, API costs, caching, context growth, latency, model routing, rate limits, concurrency, and batching using Groq.

Windows utility focused on ollama model manager, model library, and download local models for local AI workflows.

Windows utility focused on model download manager, resumable downloads, and local model queue for local AI workflows.

A skill for AI coding Agents; Free OEM service manuals -> procedure, torque table, labelled figures. 69 makes, 1960-2025. Agent skill + stdlib navigator.

Context window optimization for AI coding agents. Sandboxes tool output (98% reduction), persists session memory via SQLite + FTS5, and enforces routing across 17 platforms through MCP + hooks. Works with Claude Code, Gemini CLI, VS Code Copilot, OpenCode, Codex CLI.

Local gateway that routes coding-agent tool decisions through Jev and can attach a read-only advisor (GPT or Claude).

DSH 防退化套件｜不止打断，还会清创：把退化的思考块从模型可见历史里折叠掉，下一轮不再回传。复读分两型——A 型同一句死磕、B 型短句轮换（最难抓，实测英文 8 块全漏）。两个插件一套闭环，缺一个都会漏。

Illustrated comic-style cheat sheets for ML, LLMs, AI safety, LeetCode and Git, plus a SwiftUI study app (Panels) to read them and quiz yourself.

AI 模拟面试官：每条扣分都强制引用你回答的逐字原文（机器校验，编造引用会被重试/丢弃），四维度加权由确定性代码计算而非模型现编，评分体系设计为可经真人盲评校准 · FastAPI + Vue3 + LangGraph

YouTube transcript search with timestamps, speaker appearances, mentions, sponsors and recommendations. Official Arcmira TypeScript SDK and CLI.

A curated, bilingual list of open-source finance agents, organized by financial function and reliability. Not investment advice.

Open-source Android assistant that replaces Gemini. Summon it by gesture, bring your own API key, and it reads the screen, taps through apps and runs scripts. No backend, no telemetry.

Omo Space: real AI agents at desks inside Minecraft. Hermes, Claude Code and Codex terminals on the walls, real web pages, agent players that join your world. MIT.