Agno is a Python framework for building, orchestrating, and deploying AI agents. It provides a unified abstraction over dozens of LLM providers (OpenAI, Anthropic, Gemini, Groq, DeepSeek, etc.) and supports rich tool integration, multi-agent teams, workflow engines, RAG knowledge bases, persistent session storage, and human-in-the-loop patterns. The architecture centers on core classes: `Agent`, `Team`, `Workflow`, `Knowledge`, and `Toolkit`, all composable through a consistent configuration interface. Agents can be served via the `AgentOS` runtime, which exposes REST, MCP, A2A, Slack, and WhatsApp endpoints with JWT-based auth, scheduling, and observability. The framework includes a comprehensive evaluation suite (`AccuracyEval`, `PerformanceEval`, `ReliabilityEval`), a learning machine for persistent user profiling and memory, and an environments module for reproducible agent rollouts and scorer-based export. While primarily an application framework, it also exhibits data/research characteristics through its evaluation infrastructure, knowledge chunking/embedding pipelines, and systematic experiment tooling. The extensive cookbook (300+ examples) doubles as integration tests, demonstrating patterns ranging from simple single-agent tasks to complex distributed teams and nested workflows. Key entry points include `agno.agent.Agent`, `agno.team.Team`, `agno.workflow.Workflow`, `agno.os.AgentOS` (`agno serve`), and the `agno` CLI. The codebase is large and actively maintained, with a broad integration surface across models, vector databases, storage backends, and third-party tools. Technical drift is moderate, with some experimental model wrappers and rapidly evolving model IDs that may require updates.
jc (JSON Convert) is a mature Python library and CLI tool that converts the output of over 150 command-line tools, system files, and common data formats into structured JSON. It operates both as a piped command-line utility (`jc <command>`) and as a programmable Python module (`jc.parse('parser_name', data)`), enabling seamless integration with jq, scripting, and automation pipelines. The architecture follows a consistent parser pattern: each parser module defines an `info` metadata class and a `parse(data, raw, quiet)` function, with optional `_process()` for type coercion. Streaming parsers (suffixed `_s`) use generator-based patterns for real-time output. The library includes vendored dependencies for ASN.1, plist, EDID, and TOML parsing, and supports cross-platform execution across Linux, macOS, FreeBSD, and Windows. The primary data flow is stdin or file input → parser module → structured dict/list output (JSON or YAML). Key entry points include the CLI (`jc/cli.py`), the library API (`jc/lib.py`), and individual parser modules under `jc/parsers/`. The project maintains an extensive test suite with fixtures captured from multiple OS versions, ensuring robustness against command output variations. It also provides shell completions, a plugin system for custom parsers, and documentation generation scripts. While the codebase is well-engineered with clean abstractions, the sheer number of parsers and reliance on vendored libraries introduce moderate maintenance surface area.
AutoGPT is a full-stack SaaS platform for building, deploying, and running AI agents. It provides a visual agent builder (blocks), a copilot chat interface that can generate and orchestrate agents from natural language, a marketplace for sharing agents and skills, team/experts management with scheduling and credential stacks, and a subscription billing system with Stripe. The architecture is a multi-service backend (FastAPI, PostgreSQL, Redis) with a Next.js frontend, supporting real-time WebSocket execution events, OAuth integrations with 45+ providers, and feature flags via LaunchDarkly and PostHog. The platform is dual-licensed (Polyform Shield for platform code, MIT for classic code) and supports both hosted SaaS and self-hosted deployment. The codebase is organized into two major areas: `autogpt_platform/` (the current platform) and `classic/` (legacy AutoGPT and Forge). The platform backend uses a block-based execution system where each integration (Google, Slack, GitHub, etc.) is a self-contained block with typed I/O schemas, cost tracking, and credential management. The frontend features a rich component library with dynamic form rendering (RJSF), responsive dialog/drawer patterns, and extensive state management via TanStack Query and Zustand. Key entry points include the copilot chat UI, the agent library, marketplace, team management, and admin panels. Primary data flows involve user input through the copilot or builder, which triggers graph execution via the execution manager, with events streamed over WebSocket. Credits are deducted per block execution, and billing is handled through Stripe subscriptions with trial management. Analytics events are sent to PostHog and DataFast for product metrics. The system relies on external dependencies including PostgreSQL (with pgvector), Redis, Stripe, LaunchDarkly, PostHog, Sentry, and numerous third-party APIs for integrations.
This repository implements the HSTU (Hierarchical Sequential Transduction Unit) architecture for large-scale generative recommendation, as described in the ICML'24 paper 'Actions Speak Louder than Words'. It provides a full-stack framework spanning custom high-performance GPU kernels (CUDA, Triton), model definitions, distributed training pipelines, and MLPerf-compliant inference benchmarking. The codebase is organized into three layers: low-level ops (jagged tensors, FP8 quantization, flash attention), modular model components (STU stacks, content/action encoders, similarity modules), and production-grade training/inference scripts (DLRM-v3). A separate research subpackage includes experimental modeling variants, indexing utilities, and an LLM agent framework (cppilot) for autonomous code generation. While the primary focus is recommendation systems, the repository also serves as a showcase for advanced GPU kernel engineering and efficient sequential transformer deployment.
TinyLLaVA Factory is a modular Python framework for building, training, evaluating, and serving small-scale Large Multimodal Models (LMMs). It provides a pluggable architecture with separate components for vision towers (CLIP, SigLIP, DINOv2), language models (Phi, TinyLlama, Gemma, Qwen2, etc.), and connectors (MLP, QFormer, Resampler), all accessible via factory registries. The training pipeline supports two-stage pretraining (alignment) and finetuning with full, LoRA, or QLoRA recipes using DeepSpeed. It includes a comprehensive evaluation suite for standard benchmarks (VQA, GQA, SQA, POPE, MMMU, etc.) and a Gradio-based web demo. The codebase is well-organized and follows clean architectural patterns, though many modules are tightly coupled to the internal framework. It is actively maintained, with recent additions like a visualization tool and support for new models.
DIJI-NES is a high-performance NES (Nintendo Entertainment System) emulator designed to run on the ESP32-S3 microcontroller. It emulates the 6502 CPU, PPU graphics, and APU audio with cycle-level accuracy, supporting mappers 0-4 (NROM, MMC1, UxROM, CNROM, MMC3). The system renders 256x240 pixel frames at near-60 FPS to an ST7789 display via LovyanGFX and outputs audio through an I2S DAC (MAX98357A). It features a dual-core architecture: Core 0 handles audio generation and display DMA, while Core 1 runs the emulation loop. The firmware includes a ROM browser menu, save/load state support, volume control, and a particle boot animation. The codebase demonstrates deep NES hardware knowledge and aggressive optimization for embedded systems, including IRAM placement, lookup tables, bank caching, and frame-skip logic. While the main application is tightly coupled to the ESP32-S3 platform, the core emulation components (CPU, PPU, APU, cartridge) are well-isolated and could be extracted for reuse in other NES emulation projects.
This repository contains raygui, an immediate-mode GUI library for the raylib game programming library. It is written entirely in C as a single header file (src/raygui.h). The library provides over 25 controls (buttons, sliders, text boxes, list views, color pickers, etc.) designed specifically for building development tools, debug overlays, and editors. The architecture follows an immediate-mode pattern where controls are drawn and return their state every frame, with no retained state managed by the library, making it highly predictable and easy to integrate into existing game loops. A powerful styling system allows complete customization of colors, fonts, and metrics, with over 20 pre-built themes provided in the styles/ directory. The library also includes a built-in icon system with 256+ 16x16 icons. While tightly coupled to raylib by default, a standalone mode (RAYGUI_STANDALONE) allows it to be adapted to any graphics library by implementing a small set of backend functions. The examples directory contains several high-quality, reusable custom widgets (curve editor, property list, file dialog) that demonstrate how to extend the library. The codebase is mature, well-documented, and actively maintained, representing a high standard of C library craftsmanship within its domain.
The Open Source Pledge website serves as the public platform for the Open Source Pledge initiative, showcasing member companies that have committed to supporting open source maintainers through financial contributions. It features a member directory, blog posts about open source sustainability, aggregated job postings from member companies, resources for advocates, and historical statistics demonstrating the pledge's impact over time. The site emphasizes transparency by publishing member reports and providing clear information about how companies can join the pledge. Built with the Astro framework and integrated with Sanity CMS for blog management, the site follows a modern Jamstack architecture. Member data is stored as JSON files in the repository, while job postings are fetched from member companies' career pages and cached periodically in data/jobSet.json. Historical tracking data is maintained in data/history.csv and updated via automated workflows. The frontend consists of reusable Astro components for displaying members, testimonials, logos, and interactive elements, with data flowing from these sources through utility functions in src/members/common.ts to the presentation layer. The codebase demonstrates strong engineering practices with comprehensive TypeScript usage, Zod-based schema validation for member data, and responsive design principles. Notable reusable components include a utility module (src/util.ts) with functions like debounceLeading that correctly implements leading-edge debounce, a YouTube carousel component (src/components/YoutubeCarousel.astro) that encapsulates state management for item navigation, and an interactive SVG blob effect (src/components/Blob.astro) that showcases advanced DOM manipulation techniques for creating responsive visual effects. While minor duplication exists in some display logic, the overall structure is clean, maintainable, and exhibits several patterns worth studying for frontend development.
Generic Mapping Tools (GMT) is a mature, open-source collection of approximately 100 command-line tools for processing geographic and Cartesian data and producing high-quality PostScript illustrations, maps, and animations. The system is architected as a modular C library with a central dispatcher executable (`gmt`) that loads and runs individual modules. It supports both a classic mode (direct module calls) and a modern mode (`gmt begin/end`, `gmt subplot`) that manages sessions, figures, and panels. Key entry points include the `gmt` command, the C API (`GMT_Create_Session`, `GMT_Call_Module`), and shell completion scripts. Data flows involve reading grid, table, image, and color palette table (CPT) data from local files or remote servers (via `@` prefix), processing through modules (e.g., `grdimage`, `psxy`, `grdmath`, `grdfilter`), and outputting PostScript, grids, or other formats. External dependencies include netCDF, GDAL, FFTW, PCRE/POSIX regex, PROJ, GSHHG coastlines, and DCW country polygons. The codebase exhibits high craftsmanship with systematic option parsing, a well-defined API, extensive test coverage, and reusable infrastructure components.
Awkward Array is a Python library for manipulating nested, ragged, and mixed-type data structures that do not fit naturally into NumPy's rectangular array model. It is primarily a software library, not a data-research pipeline, though it contains research-oriented studies and benchmarks. The architecture is a Python front-end backed by compiled C++ kernels, with optional integrations for PyArrow, JAX, CuPy, Numba, CUDA, and ROOT. The core data model is built around composable layouts, forms, and types, and operations are dispatched through a high-level function decorator that separates public API from backend-specific implementations.
Lifo is a browser-based virtual machine and sandbox environment that emulates a Linux-like OS entirely in JavaScript. It provides a virtual filesystem (VFS), network stack, process management, shell, and an extensive Node.js compatibility layer that allows running real npm packages, dev servers (Vite, Expo, Express), and CLI tools inside the sandbox. The repository is a pnpm monorepo containing multiple packages: `@lifo-sh/core` (kernel, VFS, shell, node-compat), `@lifo-sh/ui` (React components for terminal, file explorer, preview), `lifo-sh` (CLI for daemon/session management), `apps/playground` (interactive web demo), `apps/desktop` (macOS-style web desktop), and many `lifo-pkg-*` command packages (git, ffmpeg, vi, etc.). The architecture is modular: the core kernel is framework-agnostic and can be embedded in Node.js or browser environments, while the UI and CLI are separate consumers. Key data flows involve user input through the shell, command execution against the VFS and network stack, and Node API calls intercepted by the compatibility layer. External dependencies include `acorn`, `browser-metro`, `ws`, and optional `esbuild`. The project is actively maintained with frequent releases, extensive tests, and a clear roadmap for expanding the sandbox capabilities.
ConcurrentSim.jl is a discrete-event process-oriented simulation framework for Julia, directly inspired by Python's SimPy. It provides a simulation engine where user-defined processes (implemented as coroutines via the ResumableFunctions.jl @resumable macro) yield events that are scheduled on a priority-ordered event heap. The framework offers a complete set of resource primitives for modeling limited-capacity congestion points: Resource (a lock-like container with capacity ≥1), Container (arbitrary capacity with continuous levels), Store (unordered collection with item filtering), QueueStore (FIFO), StackStore (LIFO), and DelayQueue (FIFO with configurable latency). Event operators (&, |) provide combinatorial AND/OR synchronization, and the framework supports both Float64 simulation time and wall-clock DateTime-based scheduling. The architecture is cleanly separated across modules: the Simulation environment manages the event heap, clock, and active process; a shared BaseEvent struct implements the callback mechanism; processes integrate with the event loop via a custom finite-state machine iterator from ResumableFunctions; resources build on a common trigger/dispatch abstraction in resources/base.jl. The codebase displays mature engineering practices: CI across Julia 1.10–1.12 plus nightly (Windows, macOS, Linux), automated static analysis via Aqua and JET, a 14-file test suite with doctests, a benchmark suite comparing old and new process implementations, and comprehensive documentation. It is the successor to SimJulia and one of the longest-lived Julia packages. The repository also contains three example notebooks, but these are legacy artifacts targeting Julia 0.6 (using deprecated APIs like put/release/srand) and should be considered historical rather than current usage guides.
Alibaba Druid is a production-grade Java library providing a high-performance JDBC connection pool, a multi-dialect SQL parser, an SQL injection firewall, and built-in monitoring/statistics. In production, it replaces or augments standard connection pools (HikariCP, DBCP) with advanced features like connection leak detection, PSCache, and web console monitoring. The SQL parser supports 30+ dialects (MySQL, PostgreSQL, Oracle, Hive, ClickHouse, etc.) and is used for SQL formatting, parameterization, and analysis. The WallFilter provides configurable security rules to block SQL injection and dangerous operations. The library integrates deeply with Spring Boot via dedicated starters for Boot 2/3/4, and exposes metrics through JMX. Architecture is organized around a filter chain pattern for intercepting JDBC calls, a visitor pattern for AST manipulation, and an SPI-based provider system for firewall rules.
Postiz is a full-stack, open-source social media scheduling and publishing platform that supports 30+ platforms (Instagram, TikTok, LinkedIn, YouTube, X, Facebook, etc.). The system is built as a pnpm monorepo with four main applications: a Next.js frontend (App Router), a NestJS backend REST API, a Temporal-based workflow orchestrator for reliable post publishing and token refresh, and a Chrome extension for cookie-based authentication. It also includes a CLI app, a public SDK (@postiz/node), and a shared libraries layer for business logic, database access (Prisma + PostgreSQL), and integrations. Key architectural patterns include an abstract social provider base class (SocialAbstract) that all platform integrations extend, a Temporal v1.8+ workflow engine with versioned posting workflows (v1.0.1 through v1.0.8), an upload factory abstracting local and Cloudflare storage, an MCP server for AI agent tools (scheduling, content generation via LangGraph), a plug/auto-reply system, and a short-link abstraction with multiple providers. The frontend uses Zustand for state management, SWR for data fetching, CopilotKit for AI textarea integration, and Mantine/Tailwind for UI. The system handles OAuth flows, subscription billing via Stripe, team collaboration, webhooks, and admin monitoring. The codebase demonstrates strong engineering with reusable abstractions for social integrations, storage, email, and short-linking, though it carries some technical debt from organic growth (multiple workflow versions, hardcoded strings, and suppressed TypeScript strictness).
go-swagger is a long-lived CLI and library for OpenAPI/Swagger 2.0 tooling, centered on code generation. The `swagger` binary (cmd/swagger) exposes commands for validate, diff, expand, flatten, mixin, serve, init, version, and generate; `swagger generate` produces Go server, client, model, operation, markdown, and CLI code from a Swagger spec. The generation pipeline loads and validates a spec through the go-openapi stack, flattens it, builds an intermediate representation of models/schemas/operations, then renders `.gotmpl` templates into Go source, finishing with a custom lightweight Go formatter and import sorting. The generated servers and clients depend on go-openapi runtime packages and can embed the spec and serve API docs. The tool also shells out to the Go toolchain at runtime, which is why the Docker image includes Go.
Golem is a production-grade, self-evolving GeoAI agent purpose-built for geospatial (GIS) workflows. It bridges natural language with professional GIS tooling (GDAL, PostGIS) and operates as a standalone application with three access modes: a terminal TUI (Bubbletea), an embedded WebUI (Vue.js), and multi-channel IM bots (Telegram, Discord, Slack, Feishu, WhatsApp, QQ, DingTalk, MaixCam). The architecture is a modular monolith in Go, centered around an agent loop that processes messages, executes tools via a guarded registry, enforces policy (strict/relaxed/off), and supports human-in-the-loop approval. It includes a message bus for decoupling, a session manager for persistent conversation history, and a context builder that dynamically composes system prompts from workspace files with cache invalidation. The system also features a learned pipeline reuse mechanism, a spatial SQL codebook with verified PostGIS patterns, and a skill telemetry system for self-improvement. While primarily an application, it incorporates data-intensive components like the geo codebook and pipeline learning, giving it a hybrid character with a strong application dominance.
The Odin Project is a Ruby on Rails 8.1 web application that delivers a free, open-source full-stack web development curriculum. It manages the entire student journey: user registration (Devise with GitHub/Google OAuth), learning paths (Foundations, Full Stack Rails, Full Stack JavaScript), course/lesson navigation, lesson content (markdown imported from GitHub via Octokit), project submissions with community voting and flagging, progress tracking with charts, and an administrative panel for moderating flags, managing team members (with role-based authorization and two-factor authentication), and viewing reports (daily sign-ups, lesson completions). The frontend uses Hotwire (Turbo + Stimulus), Tailwind CSS, ViewComponents, and esbuild, with background jobs powered by Sidekiq and cron scheduling. The architecture follows a standard Rails MVC pattern with service objects, concerns, and value objects, enhanced by a custom component library and a Tailwind form builder. It also includes Discord notification integration for daily summaries and flag alerts.
Material for MkDocs is a comprehensive documentation framework built on top of MkDocs, the Python static site generator. It provides a rich, modern theme with extensive customization options, including color palettes, fonts, icons, and layout features. The project has evolved into a full plugin ecosystem with built-in support for blogging, search (with CJK segmentation), social card generation, image optimization, privacy-aware asset proxying, multi-project builds, and offline documentation. The architecture is a hybrid of Python (MkDocs plugins, Jinja2 templates) and TypeScript/SCSS (reactive frontend components using RxJS). The theme is configured entirely through YAML, and its plugin system follows MkDocs' standard event hooks with custom priority management. In production, the system takes Markdown content and user configuration, processes it through a pipeline of plugins (blog, tags, search, social, privacy, optimize, etc.), and generates a static HTML site with a JavaScript-enhanced frontend. The frontend implements instant navigation via XHR, a full-text search engine using Lunr.js with custom tokenization, and reactive UI components for tooltips, annotations, content tabs, and more. The project is self-documenting, with its own documentation site built using the theme itself. Notably, the project has announced end-of-life (May 5, 2027) and is in maintenance mode since November 2025, with no new features planned. The successor is Zensical. Despite this, the codebase remains a high-quality reference for MkDocs theme development, reactive frontend patterns, and plugin architecture. The repository contains both source (`src/`) and generated (`material/`) directories, which poses a drift risk but also demonstrates a clear build pipeline.
Twine is a PHP string manipulation library providing an expressive, fluent object-oriented interface. It wraps a string in a `Str` object and offers methods for transformation, encoding, encryption, hashing, comparison, case conversion, and more. The library leverages PHP 8.1 features like enums for configuration and relies on the Multibyte String extension for Unicode support and OpenSSL for encryption. The architecture is a monolith single class using traits to organize functionality, with factory methods and a global helper for easy instantiation. The primary entry point is the `PHLAK\Twine\Str` class, which can be created via constructor, static `make()`, or the `str()` helper. Data flows are method chaining; most methods return new `Str` instances to maintain immutability. External dependencies are limited to PHP extensions (mbstring, openssl). The test suite is comprehensive with high coverage. The library is well-documented with a VitePress documentation site.
Django-Flags is a reusable Django application that provides a feature flag system with configurable conditions. It allows developers to toggle functionality in code, templates, and URL routing based on conditions such as boolean, user, date, request parameters, path matches, and custom conditions. Flags can be defined in Django settings or stored in the database and managed via the Django admin. The library includes decorators for views, template tags for Django and Jinja2, URL pattern helpers for conditional routing, and debugging support via Django Debug Toolbar panels. It is actively maintained, supports Django 4.2 through 6.0, and follows modern Python packaging and linting practices.