⚡ Live Prediction Demo
PulseNet learns command→output patterns, then predicts outputs of UNSEEN commands.
📊 V10.0 Benchmark Results
⚠️ All test data was never seen during training — true generalization.
Metric — Top-1 next-byte accuracy + Bits Per Byte (BPB, cross-entropy in base 2).
Datasets — English prose (70 train / 30 test), C source code (16/6), terminal commands (33/17).
Baseline — Bigram model (order-1 Markov) on identical splits for fair comparison.
Timing — Wall-clock on single CPU core (no GPU), includes training + inference.
Reproducible — Deterministic engine. Same input → same output. No random seeds.
| Benchmark | Accuracy | BPB | Speed | Memory |
|---|---|---|---|---|
| 🔤 English | 48.1% | 3.07 | 40ms | ~10.5MB |
| 💻 C Code | 66.3% | 2.23 | 25ms | ~3.1MB |
| 🤖 CMD→Result | 89.8% | 0.73 | 20ms | ~0.9MB |
| 🔁 Synthetic Patterns | 99.8% | — | <1ms | — |
⚖️ PulseNet vs Edge AI Solutions (typical deployments)
🧠 PulseNet V10
📦 TF Lite / ONNX / Edge ML
Comparison based on typical edge scenarios; values vary with model, quantization, and hardware. PulseNet targets a different niche than LLMs: real-time reflex prediction on constrained devices.
✨ Key Features
Zero Hallucination
Cannot invent false information by design. Only reproduces patterns learned from training data, or stays silent.
Hallucination = output unsupported by learned history (fabrication). Error = prediction grounded in data but incorrect. PulseNet can make errors — it cannot hallucinate.
Real-Time O(1)
Each prediction completes in constant time. No matrix multiplication, no attention heads, no transformer layers.
Online Learning
Learns in real-time while predicting. Adapts to new patterns instantly without retraining.
Fits Anywhere
5MB RAM, 70KB binary. Runs on Raspberry Pi, ESP32, or any device with a C compiler.
600× More Efficient
~0.5W on CPU vs 300W+ for GPU inference. Can run on battery or solar power.
Based on single-core CPU (~0.5W) vs typical GPU inference (300W+). Actual ratio varies by hardware.
Model Persistence
Save trained models to disk, reload instantly. Zero re-training on restart. Deploy once, persist forever.
Extended Prediction Horizon
V10.1 captures longer-range patterns for improved disambiguation on complex, structured data streams.
100% Local & Private
No cloud, no API keys, no data leaves the device. Perfect for sensitive applications.
🔧 SDK — Integrate in Minutes
PulseNet provides a minimal C API for deterministic, real-time next-byte prediction with confidence output:
Single Header
One pulsenet.h header + a compiled library. No dependencies (libc only). No runtime, no heavy toolchain.
IP Protected
Delivered as a compiled binary (e.g., libpulsenet.a / .so / .dll). Engine source code is never shipped to end users.
13 API Functions
create, destroy, reset, train, predict, probability, feed, predict_sequence, get_stats, save, load, set_license, version — that's it.
🏗️ Architecture
🚀 Applications
Medical Devices
Real-time biosignal prediction on embedded devices. ECG anomaly detection at 0.5W power.
Cybersecurity
Network traffic anomaly detection. Learns normal patterns, flags deviations in real-time.
Industrial IoT
Predictive maintenance from sensor streams. Detects pattern changes before equipment failure.
Telecommunications
Adaptive data compression on bandwidth-limited links. BPB approaching theoretical optimum.
🎯 Ideal For / Not a Fit
✅ Great Fit
- ✅ Edge / IoT / embedded devices
- ✅ Real-time pattern prediction
- ✅ Offline-first, no cloud dependency
- ✅ Deterministic, auditable outputs
- ✅ Low-power / battery-operated
- ✅ Privacy-critical environments
❌ Not a Fit
- ❌ General-purpose NLP / chatbots
- ❌ Image / video generation
- ❌ Cloud-first with unlimited budget
- ❌ Tasks requiring world knowledge
- ❌ Large language model replacement
- ❌ Creative text generation
💼 Licensing
Embedded
Per device or volume. Binary + header + examples included. Ideal for IoT products.
Contact for pricing →Server / On-Prem
Per instance. Binary + header + priority support. For server-side deployments.
Contact for pricing →Enterprise
Custom SLA, dedicated support, NDA, roadmap input. For large-scale deployments.
Contact for pricing →All tiers include: compiled binary, header file, integration examples, documentation.
❓ Frequently Asked Questions
What data types does it support?
Any byte sequence — text, code, sensor data, network packets, binary protocols. If it's bytes, PulseNet can learn patterns from it.
What about privacy / PII?
100% local. Zero data leaves the device. No cloud, no telemetry, no API calls. The engine runs entirely on your hardware.
What are the known limitations?
Performance scales with training data quality and domain specificity. Works best on structured, repetitive patterns (commands, protocols, sensor streams). Less suited for open-domain natural language.
Can I evaluate before licensing?
Yes. We offer a 7-day evaluation kit under NDA: compiled binary, header, sample datasets, and reproducible benchmark. Contact us to get started.