Browser-Native 3D Martial Arts Action Engine (WebGPU)
A high-performance browser-native 3D martial arts action engine and interactive evaluation portal built with Next.js, TypeScript, and Three.js WebGPU. Players engage in aerial dragon combat within a stylized sumi-e ink wash world, executing physical strikes governed by real-time kinematics and adaptive signal filtering.
Executes GPU-accelerated ink wash post-processing, volumetric particle splatters, and dynamic combat trails directly via TSL (Three.js Shading Language) and compute kernels, sustaining locked 60 FPS pacing without CPU draw-call bottlenecks.
Derives instantaneous 3D strike velocity (m/s) and kinetic force (N) from motion vectors with zero server roundtrip delay. Incorporates an adaptive One Euro Filter signal pipeline attenuating coordinate jitter by 88.4% while preserving sub-15ms fast-twitch strike responsiveness.
Touchless 3D Mesh Sculpting & WebGPU Spatial CAD
HandForge is a high-performance touchless 3D mesh sculpting studio and spatial CAD application running entirely client-side in standard web browsers. It fuses monocular webcam hand tracking with WebGPU WGSL compute shaders, allowing digital sculptors and 3D artists to push, pull, smooth, and sculpt high-density 3D geometry in mid-air with zero physical touch.
Utilizes WebGPU compute shaders and WGSL kernel dispatch to calculate real-time radial Hermite falloff polynomials across 125,000+ vertices at 0.38ms GPU dispatch time, bypassing the CPU bottleneck entirely to maintain a locked 60 FPS.
Tracks 21 3D hand landmarks via MediaPipe and WebAssembly. Calculates spatial palm normal quaternions and finger pinch distances for precise depth-aware brush radius scaling, dynamic displacement strength, and continuous 3D orbital mesh manipulation.
Touchless 2D Window Manager & Spatial Web OS
A high-performance touchless 2D spatial window manager running entirely in-browser. Built with MediaPipe hand landmark tracking and CSS Glassmorphism, enabling users to fluidly control an entire Web OS with zero physical contact across any camera-enabled device.
Tracks 21 3D hand keypoints in real time via MediaPipe and WebAssembly at a locked 60 FPS. Features proprietary velocity smoothing, single-finger pointer hover, and multi-finger spatial grab-and-drag window docking.
Built upon modular CSS glassmorphism, dynamic HUD telemetry graphs, interactive window docking, and client-side AI toolchains with sub-10ms input response latency.
Kinetic Browser Motion Combat Game
Shadow Boxer transforms any standard webcam into a full-body motion-sensing combat arena. Players physically dodge, weave, and deliver kinetic strikes against dynamic virtual opponents in real time.
Powered by high-speed browser computer vision, the engine tracks torso keypoints and fist trajectory vectors at 60 FPS. Punch velocities and hit-box intersections are calculated through real-time vector collision math directly in local memory.
Delivers tactile visual feedback via screen-shake shaders, particle burst emissions, and reactive sound synthesis without requiring wearable sensors or external controllers.
Client-Side Spatial Learning & AI Memory Palace
VibeLocus is an educational content creation and spatial learning architecture running 100% client-side in the browser. It parses raw documents (PDF, DOCX, TXT, Markdown) into structured hierarchical syllabi, anchors subtopics with real-time AI tutoring, and indexes knowledge representations into an in-browser Memory Palace database.
Vector search powered by local 384-dimensional embeddings (all-MiniLM-L6-v2) running via WebAssembly and WebGPU without server roundtrips. Features sub-4ms cosine similarity calculations across indexed document chunks in local memory.
Persists study notes using a structured Wing/Room/Drawer hierarchy with lossless LZString data compression (~62% reduction) and real-time quota telemetry. Orchestrates multi-provider AI tutoring with Gemini, OpenAI, Ollama, and window.ai.
60 FPS Monocular Computer Vision Engine
A breakthrough spatial perception engine that delivers true 3D perspective-correct parallax visuals on ordinary flat displays, entirely eliminating the requirement for bulky, expensive head-mounted displays.
Leveraging a Python-based computer vision core with real-time OpenCV facial landmark triangulation, the system continuously estimates user head position in 3D physical coordinates relative to the screen. Off-axis projection matrices are dynamically computed on every render frame.
Achieves persistent 60 frames per second at sub-12ms frame processing latency on standard 720p webcams, featuring adaptive Kalman filtering to smooth micro-jitters without introducing perceptual lag.
3D Spatial Hand-Tracking CAD Interface
An experimental touchless volumetric interface that liberates 3D modeling from physical mice and styluses, empowering digital creators to sculpt, translate, and orient 3D meshes directly in mid-air.
Processes 21 3D hand landmarks via MediaPipe and WebAssembly. Palm normal vectors, index-thumb pinch thresholds, and wrist flexion angles are transformed via quaternion math into direct 3D vertex operations in Three.js.
Employs Bounding Volume Hierarchy (BVH) spatial raycasting to deliver precise mid-air object selection and translation with zero perceptible delay.
Contactless Hand-Gesture Polyphonic Synthesizer
A responsive contactless acoustic synthesizer transforming standard webcams into multi-octave musical instruments through real-time fingertip trajectory analysis.
Tracks discrete fingertip coordinates across multiple octaves projected into physical space. Key-strike trigger events fire when downward Y-velocity exceeds threshold gradients, triggering dynamic ADSR polyphonic oscillators.
Engineered entirely on browser Web Audio API audio nodes without external sample latency. Features harmonic overtones, pitch-bend tracking, and dynamic polyphonic voice allocation.
Kinetic Gesture-Controlled Filter Engine
An augmented-reality mirror and digital avatar playground where visual shaders, particle systems, and aesthetic filters are commanded solely through physical gesture signifiers.
Executes parallel computer vision pipelines: 468-point dense facial mesh topology combined with continuous hand gesture classification. Custom state machines instantly switch visual shaders upon recognized gesture prompts.
Delaunay mesh triangulation and UV texture warping run 100% locally in WebAssembly. Zero video frames are transmitted to remote servers, providing total client-side privacy.
Explore our laboratory whitepapers, WebGPU compute benchmarks, and official NPM documentation.
Adaptive One Euro Filter coordinate smoothing, strike velocity (m/s) and kinetic force (N) vector derivation, LZ-String UTF-16 compression, and .hf3d binary serialization in Ink of the Forgotten Path.
Mathematical analysis of WGSL parallel dispatch, radial Hermite falloff polynomials, and zero-CPU vertex displacement at 60 FPS in HandForge.
Engineering NexaDesk: virtual viewport docking heuristics, z-index hierarchy, backdrop-filter glassmorphism, and sub-10ms gesture state machines.
Upper-body landmark trajectory calculations, instantaneous punch velocity derivation, and discrete bounding-box collisions in Shadow Boxer.
Local 384-dimensional vector embeddings (all-MiniLM-L6-v2), hierarchical IndexedDB storage with LZString compression, multi-provider LLM orchestration (useLLM), and AST security.
Computing 3D spatial palm normal vectors, adaptive Kalman filtering for pointer jitter elimination, and SIMD WASM execution on webcams.
Empirical benchmark comparing cloud streaming models against 100% in-browser WASM inference for total camera privacy and 94% latency reduction.
Designing the next generation of contactless spatial computing, low-latency edge AI, and console-grade WebGL graphics directly inside standard web browsers.
At Animatrous, our core engineering philosophy centers on 100% client-side computing. Rather than streaming raw camera video feeds to remote cloud servers—which introduces severe network latency (100ms+), bandwidth expenses, and critical privacy vulnerabilities—all computer vision models execute locally within the client's browser runtime.
By compiling complex neural networks to WebAssembly (WASM) and offloading matrix computations to WebGL shaders, our applications achieve persistent 60 frames per second on everyday commodity laptops without external GPUs.
Real-time interactive tools and kinetic web games require frame render times strictly below 16.6 milliseconds. In high-throughput JavaScript applications, heap memory allocations can trigger garbage collection pauses that destroy immersion.
Our architectures employ rigorous object pooling, typed arrays (Float32Array), and reusable geometry buffers to ensure zero mid-loop allocations. Render loops decouple sensor processing from rendering ticks, ensuring silky-smooth visuals regardless of camera frame rate variations.
Privacy is non-negotiable in spatial computing. Our tools utilize camera streams solely as transient, in-memory buffers to compute 21-point hand or facial keypoint coordinates. Video frames are immediately discarded after landmark estimation.
No image data is ever recorded, written to disk, cached, or transmitted over any network socket. Users enjoy touchless spatial interaction with absolute architectural assurance that their physical environment remains entirely confidential.
Technical specifications, browser compatibility, camera privacy standards, and integration guidelines for Animatrous web applications.
