ManimStudio
v1.1.3.0Create stunning mathematical animations with a powerful GUI application. Microsoft Store compatible — no administrator rights required.
What's Included
- ManimStudio GUI Application
- Python 3.12.7 (bundled installer)
- MiKTeX Basic (officially recommended for Manim)
Requirements
- Windows 10+ (64-bit)
- No admin privileges needed
- Internet for initial setup
Quick Install
- Run ManimStudio installer
- Auto-installs dependencies
- Follow setup wizard
- Launch and create!
Technical Specifications
Created by: Yu Yao-Hsing
License: MIT License
LaTeX Engine: MiKTeX
Format: MSIX Package
Platform: Windows 10+
ManimStudio (iOS / iPadOS)
iOS 17+ App Store Python 3.14A complete offline Python animation studio for iPad and iPhone, built on the Manim engine. Edit Python in a Monaco editor, render to MP4 via Apple VideoToolbox hardware encode, drop in LaTeX with busytex — entirely on-device.
What's Included
- Bundled Python 3.14 stack via python-ios-lib — manim, NumPy, SciPy, matplotlib, Plotly, PyAV, pycairo, pangocairo, busytex (all arm64-iphoneos)
- Monaco editor in WKWebView — full Python autocomplete, find/replace, multi-cursor, snippets, render-error gutter markers
- Live symbol-completion index built from introspection of the bundled Python packages
- Drag-drop image → auto-inserts
ImageMobjectsnippet - Pure-Swift Python formatter (⌥⌘L)
- Reuses the offlinai_shell from CodeBench / BenchCode (rebranded as "ManimStudio shell" at install)
Requirements
- iOS 17.0+ / iPadOS 17.0+
- Apple Silicon (arm64-iphoneos)
- Magic Keyboard supported
- No internet connection needed
Render Pipeline
- VideoToolbox H.264 (~5× faster than libx264)
- Cairo via pycairo compat layer
- busytex WASM for Tex / MathTex
- Output to Documents/ — Files app visible
Technical Specifications
Created by: Yu Yao-Hsing
App Store ID: 6764472686 (manimstudio)
Branch: manim_app/ios
Python: 3.14 (iOS arm64)
Editor: Monaco · Terminal: SwiftTerm + PTY
CodeBench
iPad · iPadOS · Mac Catalyst Local AI App Store name: BenchCodeA self-contained developer / scientific / AI workstation for iPad and Mac. Python 3.14, C, C++, Fortran, pdflatex, and local LLMs — all running on-device, no internet required.
Code Editor VS Code
Real Monaco editor in WKWebView. Python IntelliSense with ~70-entry signature DB, hover docs, and live resolve from the embedded Python daemon for numpy / scipy / sklearn / matplotlib / sympy completions.
Integrated Terminal PTY
SwiftTerm-backed terminal piped through a real PTY into the embedded CPython REPL. Type Python directly, or use POSIX-style builtins like ls, grep, ncdu, top, git clone.
On-Device pdflatex WASM
busytex (pdftex 1.40.25 + xetex + luatex + bibtex8) running in a hidden WKWebView. TeX Live 2023 preloaded into MEMFS plus a 23 MB overlay with PGF/TikZ/beamer, hyperref, mathtools, microtype, fontspec, CJK.
Local LLMs GGUF
llama.cpp as an XCFramework — load any Llama / Mistral / Qwen / Phi GGUF model and chat with streaming tokens. ExecuTorch backends for Apple Core ML / XNNPACK kernel-optimized PyTorch inference.
RAG & Image Gen offline
In-process sentence-embedding + vector store over user-imported text / PDF / markdown. Offline image generation via ExecuTorch-runnable diffusion models. All models stay in the app sandbox — no tokens leave the device.
Workspace System multi-tab
iOS document browser with multiple concurrent workspaces. Auto-save debounced ~600 ms after every keystroke + on run, tab-switch, view-disappear, and app-backgrounding. Tombstone system records deleted files so seeders don't resurrect them.
Metal GPU PyTorch
Monkey-patched torch.matmul / mm / bmm / addmm, F.linear, and SDPA onto Apple's MPSMatrixMultiplication via ctypes → Swift @_cdecl. 2–10× speedup at training sizes. fp32/fp16 native, bf16 via fp32 cast. Autograd-aware via torch.autograd.Function. Auto-installed via sitecustomize.py — zero user code changes.
UI Polish
Live RAM sparkline in the top-right tab bar, command palette (⌘P), editor status bar, tab slide animation, drag-to-select in the terminal (via SwiftTerm doubleTap), and a real Stop button that interrupts the running task.
Training Toolkit
Opt-in _cb_training.py (~12 KB pure Python) ships OOMGuard, MemoryProfiler, KVCache, and TrainingMonitor. Plus a pure-Python safetensors reader (mmap + struct + torch.frombuffer) so transformers.from_pretrained() works without Rust crates.
LoRA Fine-tuning NEW
Train a LoRA adapter on a GGUF base model in-place via llama.cpp's Metal backward kernels (LlamaFinetuner.swift). Closes the train→deploy loop: fine-tune with PyTorch + HF Trainer (Metal bridge for speed), then convert the .pt to a GGUF adapter with a pure-Python writer for fast Metal inference.
JavaScript REPL NEW
js shell builtin backed by Apple's JavaScriptCore (zero bundle cost) — persistent Node-style REPL globals, js -e one-liners, and js script.js file execution, sitting alongside the python builtin.
Editor Performance NEW
IntelliSense engine moved from a 3 s blocking loop to DispatchSource event watch (no held threads), Monaco file opens batched 3 JS round-trips → 1 (~100 ms saved per open), and C/C++/Fortran language providers lazy-register only on first non-Python file.
AI Assist Chat NEW
ChatGPT-style chat panel wired to the local LLM — streaming replies that render compiled LaTeX inline, code-aware assists into the editor, and a one-tap scratch file when there's no target. All on-device; no tokens leave the iPad.
Visual Debugger NEW
Step through Python with a visual debugger wired into Monaco, data Quick Look for inspecting variables / arrays / frames, plus a redesigned Libraries tab, richer System Info, and a hidden debug panel with a native-lib test probe.
Languages — all share Monaco + IntelliSense + auto-save
Built On
- python-ios-lib — runtime layer
- SwiftTerm — xterm-compatible PTY
- Monaco Editor — VS Code in WebView
- busytex — TeX Live 2023 in WASM
- llama.cpp + ExecuTorch — local AI
Bundle Composition
- ~1 GB installed app size
- 791 MB Frameworks (Python, llama, ExecuTorch)
- 484 MB Python site-packages
- 254 MB Swift source + LaTeX data
Technical Specifications
Created by: Yu Yao-Hsing
Platforms: iPad · iPadOS · Mac Catalyst
Runtime: Python 3.14 + four interpreters
Editor: Monaco (VS Code) in WKWebView
Status: Active Development
python-ios-lib
Python 3.14 30+ LibrariesFull Python 3.14 runtime for iOS/iPadOS with 30+ offline libraries — including native PyTorch, HuggingFace transformers, and Rust tokenizers. Train on-device. No JIT, App Store safe. Powers CodeBench. Its Metal-accelerated pieces also ship as standalone repos: cairometal and torchmetal.
Machine Learning 4
| Library | Ver |
|---|---|
| PyTorch BETA | 2.1.2 |
| transformers BETA | 4.41.2 |
| tokenizers BETA | 0.19.1 |
| scikit-learn | 40 modules |
Native PyTorch (95/95 asserts) + Metal GPU bridge (2–10× speedup) + HuggingFace transformers & Rust tokenizers — train on-device
Visualization 3
| Library | Ver |
|---|---|
| matplotlib | 3.9.0 |
| Plotly | 6.6.0 |
| manim BETA | 0.20.1 |
145+ mobjects, 73 animations, interactive Plotly HTML charts
Media & Rendering 5
| Library | Type |
|---|---|
| PyAV | 17 C exts |
| FFmpeg | 7 dylibs |
| Cairo + Pango | Native iOS |
| Pillow | 12.2.0 |
| offlinai_latex | SwiftMath + 33MB texmf |
Data & Web 6
| Library | Purpose |
|---|---|
| PyWebView NEW | v5.4.0 — iOS WKWebView |
| requests | HTTP client |
| BeautifulSoup4 | HTML parse |
| NetworkX | Graphs |
| jsonschema | Validation |
| PyYAML | YAML |
System & Graphics 7
| Library | Ver |
|---|---|
| psutil | 5.9.8 |
| watchdog | 4.0.0 |
| moderngl | 5.12.0 |
| svgelements | 1.9.6 |
| decorator NEW | 5.1.1 |
| safetensors | HF deps |
| huggingface-hub | HF deps |
Monaco IntelliSense ~70 sigs
Real VS Code editor in WKWebView with Python signature help, hover docs, and live resolve from Python for numpy / scipy / sklearn / matplotlib / sympy.
Expanded LaTeX Bundle 33 MB
Full Latin Modern Type 1 fonts, expl3 (1.3 MB), firstaid, hyphenation, unicode-data, pdftex.map. Math-mode rendering via SwiftMath is unlimited.
Shell Builtins POSIX-y
ncdu with arrow-key navigation, top with Apple-chip detection, git clone via zipball fetch, universal --help interception.
Auto-save ~600 ms
Edits persist to disk on every keystroke (debounced) plus on run, tab-switch, view-disappear, and app-backgrounding. No more lost edits on reopen.
Tombstone System .offlinai_deleted
Files deleted via UI / shell rm / ncdu's d are recorded so starter-script seeders (pip_demo.py, torch_test_all.py) won't re-create them.
Test Coverage 24/24
Full integration test for the transformers stack passes 24/24. PyTorch ships 95/95 correctness asserts. First public iOS builds for PyTorch + tokenizers.
Metal GPU Bridge
_torch_metal_bridge.py patches torch matmul / mm / bmm / addmm, F.linear, and SDPA onto Apple's MPSMatrixMultiplication. 2–10× speedup at training sizes. fp32/fp16 native, bf16 via fp32 cast. Autograd-aware, auto-installed via sitecustomize.py.
Git via dulwich
Real git push / pull / commit via auto-installed dulwich. Plus pywebview FileDialog enum and faulthandler that writes C-level crashes to ~/Documents/log.txt.
CairoMetal NEW
pycairo-compatible 2D vector graphics rendered on the Apple GPU — 246 cm_* functions, stencil-then-cover fills, all 28 compositing operators, IOSurface-backed MTLTexture, pixel-diffed vs real cairo. Cairo has no other Metal backend — to our knowledge this is novel. Standalone repo: cairometal — now pip install cairometal on PyPI (macOS arm64 wheel).
TorchMetal NEW
App-Store-safe per-op Metal/MPS routing for PyTorch inference — matmul, linear, softmax, layer_norm, gelu, attention — size + dtype gated with bit-exact CPU fallback. fp16 matmul is often hundreds of times faster than CPU. Public MPS only, no private symbols. Standalone repo: torchmetal.
Manim 4K/8K NEW
Memory-safe high-res rendering up to 8K (7680×4320) without tripping iOS jetsam — bounded GIF buffer, frames stream to mp4 instead of accumulating, resolution-tiered RAM pre-flight, h264_videotoolbox encode. Quality selectable 480p→8K.
JavaScript REPL NEW
js shell builtin backed by Apple's JavaScriptCore — zero bundle cost. Persistent REPL globals (Node-style), js -e "code" one-liners, and js script.js file execution. See js-engine.md.
Requirements
- iOS 17.0+ / iPadOS 17.0+
- Python 3.14 (BeeWare)
- NumPy 2.x (arm64 wheel)
- Works completely offline
Quick Start
- Copy library folders into
site-packages/ - Add interpreter .c/.h to Xcode
- Import and use
- No JIT needed!
Technical Specifications
Created by: Yu Yao-Hsing
License: MIT
Runtime: Python 3.14 on iOS arm64
Libraries: 30+ (native + pure Python)
Platform: iOS 17+ / iPadOS 17+
Bootbox
iPad · iPadOS On-Device Emulation No JailbreakA boot manager for your iPad. Bootbox is a native iOS app that boots real operating systems — 64-bit Linux (x86-64 and ARM64), classic Windows, graphical desktops — inside a WebKit-hosted emulation stack, with real internet, real multi-core, files in the iOS Files app, and one-tap downloads for every guest image. Everything runs on-device: no servers, no streaming, no jailbreak.
64-bit Linux & ARM64 QEMU-Wasm
Genuine aarch64 Alpine (uname -m really reports aarch64) alongside a dual-core x86-64 console guest with Python 3.12, pip/uv, Wine, and real internet.
Classic Windows v86
Windows 98 SE and Windows 2000 Pro boot natively to the desktop via v86, with working internet on Windows 2000 through the NDIS relay.
Real Multi-core up to 8
The x86-64 guest boots dual-core by default with genuine parallel execution (measured 2.0× wall-clock speedup). A cores selector (1/2/4/6/8) sits in the toolbar and UEFI Setup.
4× Lower Idle Power
A patched engine sleeps on a real futex between polls instead of busy-spinning. Measured: idle ~120% → ~32% host CPU, boot time 30s → ~20s, cooler iPad, longer battery.
Python That Works uv
pip install backed by uv with ~24 packages pre-installed and 10-minute download timeouts for large wheels. A baked-in constraints file resolves NumPy to the 1.x line for correctness on the emulated CPU.
Files App Integration iSH-style
A classic NSFileProviderExtension exposes the Bootbox folder in the iOS Files app. Materialized files are hard links — drop in .iso/.img, pull guest exports out, zero-copy.
Real Internet gVisor
Iosnet.xcframework embeds a gVisor userspace TCP/IP stack — the guest's virtio NIC frames travel over a local WebSocket into native sockets. pip, apk, wget, DNS all work, fully on-device.
Your Own Images
Import any .iso/.img via the Files app; paired save-states let a custom install resume in about a second instead of rebooting.
A Boot Experience
GRUB-style boot menu with per-system notes, a classic Aptio-blue UEFI Setup screen, then serial console on the left and GUI on the right — keyboard routing follows your last tap.
Built On
- QEMU-Wasm — x86-64 (SMP/ACPI) & ARM64 engines
- v86 — fast 32-bit x86 (Win98 / 2000, i686)
- Boxedwine — Wine in the browser
- noVNC — graphical guest output
- Iosnet.xcframework — gVisor netstack (gomobile)
What It Boots
- 64-bit Linux + Python & Wine (x86-64, 1–8 cores)
- 64-bit Linux Desktop (twm + terminal + browser)
- 64-bit Linux — genuine ARM64 Alpine
- Windows 98 SE & Windows 2000 Pro
- Any imported
.iso/.img
Technical Specifications
Created by: Yu Yao-Hsing
Platform: iPad · iPadOS (Swift host app + WKWebView)
Networking: gVisor userspace TCP/IP via Iosnet.xcframework
App size: ~60 MB · guest images download on demand, once, and are cached
Status: Active development — build from source (Xcode 16+, Apple developer team)
Generalized Covariance Matrix
ESD Analysis ToolAdvanced eigenvalue spectral distribution analysis for generalized covariance matrices. A powerful research tool for mathematicians and data scientists!
Key Capabilities
- Analyze eigenvalue distributions
- Visualize spectral densities
- Explore random matrix theory
- Export data for further analysis
Perfect for: Statistical analysis, ML research, quantum mechanics, and financial modeling
System Requirements
- Windows 10+ (64-bit)
- Works completely offline
- One-click installation
Get Started
- Install from Microsoft Store
- Launch the application
- Input your matrix data
- Analyze & visualize!
Technical Specifications
Created by: Yu Yao-Hsing
Category: Research & Education
Platform: Windows Store
Type: Desktop App
Status: Active Development
EigenDenoise
macOS 14+ Metal GPUNative macOS image denoiser using random matrix theory — the macOS counterpart to Generalized Covariance Matrix. Pure Swift, Metal-accelerated, App Store sandbox-ready.
Key Capabilities
- Side-by-side RMT denoisers — MP thresholding + generalized-covariance oracle (β·δ_a + (1−β)·δ_1)
- Differential evolution to find best (a, β) against a clean reference
- Three interactive spectral tabs — Eigenvalue density, Im(s) vs z, Roots vs β
- Curated dataset gallery: ORL Faces (400 PGMs), CBSD68 (68 colour), Brain MRI (3,264 T1 slices)
- Custom URL lists with per-image checkboxes & destination preview
System Requirements
- macOS 14 (Sonoma) or later
- Apple Silicon (arm64) — Rosetta supported
- Metal GPU acceleration (LAPACK fallback)
- App Sandbox enabled
Tech Stack
- Pure Swift 5.9+ / SwiftUI / Charts
MPSMatrixMultiplicationfor Gram step- LAPACK
dsyevdfor eigen-decomposition - Zero external Swift package deps
Mathematics
For X ∈ ℝ^(p×n) with i.i.d. N(0,1) entries, S_n = (1/n)XXᵀ and T_n = diag(t₁,…,t_p), the generalized sample covariance B_n = S_n T_n. Its Stieltjes transform satisfies a cubic:
a·z·s³ + (a(z − y + 1) + z)·s² + (a + z − y + 1 − y·β·(a − 1))·s + 1 = 0
Limiting density f_{y,H}(z) follows from Cardano's depressed-cubic root; support recovered via bulk-edge function and quartic discriminant for one-interval (Cases 1, 3) vs two-interval (Case 2) topology.
Technical Specifications
Created by: Yao-Hsing Yu
License: MIT
Platform: macOS 14+ (Apple Silicon)
Type: Native macOS App (App Store ready)
Status: Active Development
GPS-location-app
iOS 17+ App Store Apple Watch CarPlayPrecision GPS workout tracker for iPhone & Apple Watch. Kalman-filtered location, HealthKit sync, Live Activities, CarPlay, and route analytics — all in a clean SwiftUI interface.
Key Capabilities
- Record workouts with high-precision GPS, Kalman-filtered for noisy signals
- Continuous tracking from iPhone + Apple Watch — start one, finish on the other
- Two-way HealthKit sync — workouts appear in the iOS Fitness / Health apps
- Live Activity / Dynamic Island shows current distance, pace, and duration
- CarPlay support for in-car activity display
- Per-workout route map with split / pace / elevation analytics
- All workout data stays on your device — no servers, no telemetry
Requirements
- iOS 17.0+ / watchOS 10+
- Magic Keyboard supported
- Location, Motion & Fitness, HealthKit permissions
- Works fully offline
Get Started
- Install from the App Store
- Grant location & HealthKit permissions
- Pair with Apple Watch (optional)
- Tap Start to record a workout
Technical Specifications
Created by: Yu Yao-Hsing
App Store ID: 6764729098
Frameworks: SwiftUI · CoreLocation · HealthKit · ActivityKit · CarPlay
Stack: Pure Swift / SwiftUI
Status: Live on App Store (approved May 18, 2026)
Apps & Visualizations
WhisperKit NEW
On-device speech transcription for iPhone & iPad — Whisper models running entirely on Apple's Neural Engine. Record, import audio/video, or pick from Photos; transcripts never leave your device.
Privacy policy
NeonScribe
A Python-based creative writing and text processing tool with a modern interface.
View projectUtilities
Sound Transfer
TCP-based sound transfer tool for streaming audio between devices over a network.
View projectGoogle Drive Download
A Python tool for downloading files from Google Drive with ease and automation.
View project