kizzasi

Autoregressive General-Purpose Signal Predictor (AGSP)

Neuro-symbolic architecture for continuous signal streams with state space models and constraint enforcement.

Overview

Kizzasi (兆し - "sign/omen") is a Rust-native system for predicting continuous signal streams using state space models. Treats all modalities (audio, sensors, video, control signals) as equivalent signal streams.

Key Features

Unified Interface: Single API for all model architectures
O(1) Inference: Constant-time per-step prediction for streaming
Constraint Enforcement: Safety guardrails via TensorLogic integration
Real-Time I/O: MQTT, audio, sensors, video streams
Production Ready: 124 tests, zero warnings, full documentation
Async/Streaming: Tokio-based async prediction pipelines
Model Versioning: A/B testing, hot-swapping, canary deployments

Quick Start

use kizzasi::{Kizzasi, ModelType};

// Create predictor with Mamba2 model
let predictor = Kizzasi::builder()
    .model_type(ModelType::Mamba2)
    .input_dim(32)
    .output_dim(32)
    .hidden_dim(64)
    .build()?;

// Single-step prediction
let input = vec![0.1, 0.2, 0.3, /* ... */];
let output = predictor.step(input)?;

// Multi-step rollout
let predictions = predictor.predict_n(input, 100)?;

// With safety guardrails
use kizzasi::prelude::*;
let constrained = predictor
    .with_guardrails(my_constraints)
    .step(input)?;

Architecture

kizzasi/
├── kizzasi-core      # SSM engine, SIMD, parallel scan
├── kizzasi-model     # Mamba, RWKV, S4, Transformer
├── kizzasi-tokenizer # VQ-VAE, quantization, compression
├── kizzasi-inference # Pipeline, sampling, streaming
├── kizzasi-logic     # Constraints, optimization, safety
├── kizzasi-io        # MQTT, audio, sensors, video
└── kizzasi           # Unified facade (this crate)

Use Cases

Robotics Control: Real-time motor control with safety constraints
Anomaly Detection: Learn normal patterns, detect deviations
Audio Processing: Next-sample prediction, streaming synthesis
Sensor Fusion: Multi-modal signal integration
Video Prediction: Frame-to-frame prediction with temporal coherence

Presets

// Audio processing (16kHz, mel features)
let audio_predictor = Kizzasi::audio_preset(80)?;

// Robotics control (6-DOF arm)
let robot = Kizzasi::robotics_preset(6)?;

// Sensor streams (IoT)
let sensor = Kizzasi::sensor_preset(16)?;

// Lightweight (edge devices)
let edge = Kizzasi::lightweight_preset(32)?;

Performance

Mamba2 latency: <100μs per step
Throughput: 320K predictions/sec (distributed)
Memory: <50MB for typical models
Zero-copy operations where possible

Documentation

Citation

If you use Kizzasi in research, please cite:

@software{kizzasi2024,
  title = {Kizzasi: Autoregressive General-Purpose Signal Predictor},
  author = {COOLJAPAN Team},
  year = {2024},
  url = {https://github.com/cool-japan/kizzasi}
}

License

Licensed under either of Apache License, Version 2.0 or MIT license at your option.

Contributing

See CONTRIBUTING.md

kizzasi 0.1.0