Deep Orbit · 11 MODULES
MODULE MAP
11 modules between your frontend and a deployed model. Pick up where you left off.
The crew
- Commander MirzaMentor
- Dr. FarahAstrophysicist
- Chief NazariSkeptic
- Ensign DemirCommunications
- ARIAShip AI
M01Foundations: Tensors & DataACTIVEBoard the Archimedes and learn to see data the way neural networks do. You'll master tensors, shapes, and operations (the building blocks of every ML model) using concepts you already know from frontend development.
0/5
Board the Archimedes and learn to see data the way neural networks do. You'll master tensors, shapes, and operations (the building blocks of every ML model) using concepts you already know from frontend development.
- ○Archimedes, Bridge · 18 minIntroduction to Tensors
“The Archimedes receives a mysterious data stream from deep space.”
TypedArray → Tensorbeginner1 bridge - ○Archimedes, Lab · 18 minUnderstanding Tensor Shapes
“Dr. Farah analyzes the dimensional structure of the incoming data.”
Array.length → Tensor.shapebeginner1 bridge - ○Archimedes, Bridge · 20 minBasic Tensor Operations
“Demir's first attempt at combining the feeds just crashed the console. Nazari wants it fixed, at sensor speed.”
Array.map() → tf.add()beginner1 bridge - ○Archimedes, Engineering · 24 minLoading and Preparing Data
“Ensign Demir sets up the data pipeline for the analysis system.”
beginner - ○Archimedes, Bridge · 27 minData Normalization
“Commander Mirza orders all sensor data normalized before the deep scan begins.”
Scroll progress calculation → Min-Max Normalizationbeginner1 bridge
M02Neural Networks: Layers & ArchitectureBuild your first neural network from scratch. You'll learn how neurons, layers, and activations work, translating frontend component composition into network architecture.
0/5
Build your first neural network from scratch. You'll learn how neurons, layers, and activations work, translating frontend component composition into network architecture.
M03Training: Loss, Gradients & OptimizationTrain the DEEP SCAN signal classifier. You'll master loss functions, gradient descent, and the training loop: the engine that turns data into intelligence.
0/5
Train the DEEP SCAN signal classifier. You'll master loss functions, gradient descent, and the training loop: the engine that turns data into intelligence.
M04Browser ML RuntimesTake your trained model out of the lab and into the browser. Nazari
challenges Demir to deploy the DEEP SCAN classifier on the ship's
browser-based dashboard, with real performance constraints. You'll
master TensorFlow.js setup, model loading, real-time inference, and
compare runtimes like a frontend engineer compares frameworks.
0/6
Take your trained model out of the lab and into the browser. Nazari challenges Demir to deploy the DEEP SCAN classifier on the ship's browser-based dashboard, with real performance constraints. You'll master TensorFlow.js setup, model loading, real-time inference, and compare runtimes like a frontend engineer compares frameworks.
M05Signal ProcessingDr. Farah discovers hidden structure in the signals that only appears in the
frequency domain. You'll master the Fourier Transform, spectrograms, and
windowing, transforming raw time-series data into rich features that boost
DEEP SCAN's accuracy from 78% to 87%.
0/5
Dr. Farah discovers hidden structure in the signals that only appears in the frequency domain. You'll master the Fourier Transform, spectrograms, and windowing, transforming raw time-series data into rich features that boost DEEP SCAN's accuracy from 78% to 87%.
M06Recurrent NetworksThe signals show temporal patterns: a rhythm spanning dozens of time steps.
Feedforward networks see snapshots; recurrent networks see sequences. You'll
build RNNs, LSTMs, and GRUs that remember past inputs and predict what comes
next, using concepts from Redux reducers and localStorage caching.
0/5
The signals show temporal patterns: a rhythm spanning dozens of time steps. Feedforward networks see snapshots; recurrent networks see sequences. You'll build RNNs, LSTMs, and GRUs that remember past inputs and predict what comes next, using concepts from Redux reducers and localStorage caching.
M07Anomaly DetectionDEEP SCAN's anomaly detector fires on a signal that matches no known category.
The reconstruction error is off the charts. You'll build autoencoders that learn
what "normal" looks like, then flag anything that deviates, using the same
intuition as git diff and error boundaries.
0/5
DEEP SCAN's anomaly detector fires on a signal that matches no known category. The reconstruction error is off the charts. You'll build autoencoders that learn what "normal" looks like, then flag anything that deviates, using the same intuition as git diff and error boundaries.
M08Advanced Time-SeriesThe crew attempts to predict the anomalous signal's next occurrence. You'll
build attention mechanisms, multi-step forecasts, Kalman filters, and
confidence-weighted ensembles: advanced tools that reveal the signal's
timing is changing.
0/5
The crew attempts to predict the anomalous signal's next occurrence. You'll build attention mechanisms, multi-step forecasts, Kalman filters, and confidence-weighted ensembles: advanced tools that reveal the signal's timing is changing.
M09LLM IntegrationMirza needs a comprehensive report for Earth Command: 47 days of data,
12,000 classified signals, and one anomaly that could change everything.
You'll integrate language models into DEEP SCAN: prompt engineering, RAG
over signal logs, and multi-model orchestration.
0/5
Mirza needs a comprehensive report for Earth Command: 47 days of data, 12,000 classified signals, and one anomaly that could change everything. You'll integrate language models into DEEP SCAN: prompt engineering, RAG over signal logs, and multi-model orchestration.
M10Optimization & EdgeNazari's module, and a penance. At 0247 the 340 MB model crashed the
512 MB bridge dashboard during a predicted pulse window; the capture was
lost and Earth's relay slot closed without it. You'll master quantization,
pruning, caching, and WebGPU acceleration to shrink the model to 42 MB
with 2.4% accuracy loss, measured and refused rounding, so it never
happens again.
0/5
Nazari's module, and a penance. At 0247 the 340 MB model crashed the 512 MB bridge dashboard during a predicted pulse window; the capture was lost and Earth's relay slot closed without it. You'll master quantization, pruning, caching, and WebGPU acceleration to shrink the model to 42 MB with 2.4% accuracy loss, measured and refused rounding, so it never happens again.
M11Capstone: Deploy DEEP SCANDeploy the complete DEEP SCAN signal classification system. System architecture, CI/CD pipeline, monitoring, documentation, and launch: everything the crew has built comes together.
0/5
Deploy the complete DEEP SCAN signal classification system. System architecture, CI/CD pipeline, monitoring, documentation, and launch: everything the crew has built comes together.