Architectures, ready to open
48 architectures kept as typed graphs, not pictures: 2211 layers in total, every tensor shape propagated, 36 of them clean under all 43 single-graph checks. Open one on the canvas and it is yours to edit, verify and export.
Audio
Biosignal
Computer Vision
Generative
Multimodal
NLP/LLM
Recommendation
Reinforcement Learning
Time-series
index.json
Audio
Biosignal
🧠 EEGNet
Compact CNN for EEG/BCI — depthwise + separable convs make it 10× lighter than standard CNNs
🧠 EEG Conformer
Conv stem + Transformer encoder — SOTA for high-channel motor imagery EEG
Computer Vision
🖼️ Simple CNN
Simple Convolutional Neural Network for image classification
🔗 ResNet Block
ResNet residual block with skip connections
🩻 U-Net
Encoder-decoder with skip connections — Ronneberger et al
👁️ ViT-B/16
Vision Transformer — patch embedding stem + 1 encoder block
🪟 Swin-Tiny
Hierarchical vision transformer — shifted-window attention builds a feature pyramid for dense prediction
🧩 I-JEPA (Joint-Embedding Predictive Architecture)
CVPR 2023 - prediction in representation space rather than pixel space
Generative
🎨 Diffusion UNet
Stable-Diffusion-style noise predictor — latent UNet with cross-attention to a text embedding
🌀 DiT-XL/2
Diffusion Transformer — replaces the UNet denoiser with a ViT backbone conditioned on timestep + class via adaLN-Zero
🎞️ Latent Video Diffusion Transformer
The shape every 2024-2025 video generator shares: a causal 3D VAE compresses the clip, the latent is patchified into space-time tokens, a DiT stack denoises them
Multimodal
🔗 CLIP ViT-B/32
Dual-encoder contrastive model — a ViT image tower and a Transformer text tower projected into a shared embedding space
👁️ LLaVA-1.5
Vision-language model — CLIP image encoder + MLP projector feed visual tokens into a LLaMA decoder
🧱 Qwen2-VL (Native-Resolution VLM)
Wang et al
🔗 ColBERT (Late Interaction Retrieval)
Khattab and Zaharia 2020 - the document keeps its tokens until scoring time
🪆 Matryoshka Text Embedder
Kusupati et al
NLP/LLM
🤖 Transformer Block
Transformer encoder block
📖 BERT Base
BERT-Base encoder — bidirectional MHA
🧠 GPT-2
GPT-2 Small — causal transformer block
🦙 LLaMA-3 Block
LLaMA-3 decoder block — GQA
🔀 Mixtral MoE Block
Mixtral decoder block — GQA + Sparse MoE
🔁 T5 Small
T5 encoder-decoder — bidirectional encoder + masked decoder with cross-attention
🔄 Simple RNN
Simple Recurrent Neural Network for sequence processing
🐍 Mamba SSM Block
Mamba State Space Model — selective SSM + causal conv gating, no attention
φ Phi-3 Mini Block
Phi-3 Mini 3
🐋 DeepSeek-V3
671B MoE LLM — Multi-head Latent Attention
🦙 Llama-4 Scout
109B natively-multimodal MoE LLM — interleaved dense/MoE layers
🧬 Jamba
Hybrid SSM-Transformer-MoE — interleaves Mamba, attention, and MoE blocks in one stack
🟣 Qwen3-8B
Modern dense decoder LLM — GQA with QK-RMSNorm on the query/key projections for training stability
Recommendation
🗼 Two-Tower
User+Item dual encoder for retrieval — embeddings → MLP per side → dot product score
📐 Wide & Deep
Memorization
🛒 DLRM
Meta's Deep Learning Recommendation Model — bottom MLP for dense, embedding for sparse, feature interaction, top MLP
🤝 NeuMF (Fused GMF + MLP)
He et al
🕸 GraphSAGE Recommender
Inductive node embeddings via neighbor sampling + aggregation — for graph-based recommenders
🧮 Neural Collaborative Filtering
He et al
💡 LightGCN
He et al
🧾 Behavior Sequence Transformer
Alibaba 2019 BST — user behavior sequence + target → Transformer encoder + concat with user features → MLP → CTR
🕒 SLi-Rec
Yu et al
🎯 Deep Interest Network (DIN)
Alibaba 2018 - the candidate item is the attention QUERY over the behaviour sequence, so the user embedding is computed per candidate instead of once
🔎 Search-based Interest Model (SIM)
Alibaba 2020 - a lifelong sequence
🧭 Multi-Interest Network (MIND)
Alibaba 2019 - dynamic routing turns one behaviour sequence into K interest vectors, so a user is K points in item space rather than one
📈 HSTU (Generative Recommender)
Meta 2024 - recommendation as sequential transduction over one interleaved item+action stream, with pointwise aggregated attention instead of softmax
🐯 TIGER (Semantic ID Generative Retrieval)
NeurIPS 2023 - retrieval as sequence generation: the model DECODES the next item id code by code, so the catalogue never appears as a softmax over millions of rows
💠 RQ-VAE Semantic ID Tokenizer
The tokenizer that turns an item content embedding into a short tuple of discrete codes, giving items a coarse-to-fine hierarchy a decoder can generate
Reinforcement Learning
Time-series
📈 PatchTST
Channel-independent patching + Transformer for multivariate time-series
📈 1D CNN + LSTM
Conv1D + LSTM baseline for ECG/PPG/IMU and other long-form physio signals
For agents
Every architecture here is a graph an agent can fetch and edit. index.json lists all 48 with the graph URL, the parameter count and the verifier's verdict; each has a markdown twin. Send an edited graph to POST /api/v1/check and the same checks grade it, or run the MCP server to keep them in the loop while the agent works.