Goainglys - Native Go ML Platform

A comprehensive machine learning platform written entirely in pure Go with zero external dependencies. Implements 15+ distinct ML components from tensors to production APIs.

Architecture

┌─────────────────────────────────────────────────────────────────────────────┐
│                              Goainglys ML Platform                           │
├─────────────────────────────────────────────────────────────────────────────┤
│                                                                              │
│  ┌──────────────────────────┐    ┌───────────────────────────────────────┐  │
│  │      Core Foundation     │    │           MLOps & Tooling             │  │
│  ├──────────────────────────┤    ├───────────────────────────────────────┤  │
│  │ tensor/     Tensor ops   │    │ portal/       Unified web platform    │  │
│  │ nn/         NN layers    │    │ dashboard/    Training metrics        │  │
│  │ transformer/ Full model  │    │ model_registry/ Version control      │  │
│  └──────────────────────────┘    │ marketplace/  MCP apps marketplace    │  │
│                                  └───────────────────────────────────────┘  │
│  ┌──────────────────────────┐    ┌───────────────────────────────────────┐  │
│  │     Specialized ML       │    │             AI Agents                │  │
│  ├──────────────────────────┤    ├───────────────────────────────────────┤  │
│  │ transformers/ GPT/BERT   │    │ agents/       ReAct, Plan-Execute    │  │
│  │ vector_db/   HNSW index  │    │ self_improving_agents/               │  │
│  │ asr/         Speech      │    │            Recursive improvement     │  │
│  │ finetune/    LoRA tuning │    └───────────────────────────────────────┘  │
│  └──────────────────────────┘                                               │
│                                                                              │
│  ┌──────────────────────────────────────────────────────────────────────┐   │
│  │                      Advanced Research                               │   │
│  ├──────────────────────────────────────────────────────────────────────┤   │
│  │ Mixture of Experts │ Diffusion Models │ RL Agents │ Graph Neural Nets│   │
│  └──────────────────────────────────────────────────────────────────────┘   │
│                                                                              │
└─────────────────────────────────────────────────────────────────────────────┘

Projects

Project	Description	Status
tensor/	Core tensor operations with gradient tracking	Complete
nn/	Unified neural network layers and optimizers	Complete
transformer/	Full Transformer (Vaswani et al., 2017) with backprop	Complete
transformers/	Native GPT-2 and BERT implementations	Complete
vector_db/	HNSW-based vector similarity search	Complete
asr/	Speech recognition with MFCC & LSTM	Complete
finetune/	Fine-tuning with LoRA, AdamW, Lion optimizers	Complete
agents/framework/	ReAct, Plan-Execute, Supervisor agents	Complete
self_improving_agents/	Recursive self-improvement system	Complete
Advanced-Research/	MoE, Diffusion, RL, GraphNN implementations	Complete
rag_eval/	RAG/LLM evaluation toolkit	Complete
dashboard/	Real-time training metrics visualization	Complete
model_registry/	Model versioning and staging system	Complete
marketplace/	MCP apps marketplace with REST API	Complete
portal/	Unified web platform (all services combined)	Complete
benchmarks/	Performance benchmark suites	Complete

Quick Start

# Transformer demo
cd /home/runner/workspace && GOWORK=off go run main.go

# Vector database
cd vector_db && GOWORK=off go run .

# ASR pipeline
cd asr && GOWORK=off go run cmd/main.go

# Agent framework
cd agents/framework && GOWORK=off go run main.go

# Fine-tuning
cd finetune && GOWORK=off go run main.go

# GPT/BERT
cd transformers && GOWORK=off go run main.go

# Web services
cd dashboard && go run main.go        # :8080
cd model_registry && go run main.go   # :8081
cd marketplace && go run main.go      # :8082
cd portal && go run main.go           # :8083 (unified)

Performance

Component	Metric	Value
Transformer	Training	~2.5s/epoch
Transformer	Backprop	~8s/epoch
Vector DB	Insert	10k vectors/s
Vector DB	Search	~20µs
ASR	Speed	75x realtime
RAG Eval	Throughput	10k evals/s

Core Components

Tensor Operations (tensor/)

type Tensor struct {
    Data    []float64
    Shape   []int
    Grad    []float64
    ReqGrad bool
}

Transformer Config (transformer/)

type Config struct {
    ModelDim  int
    NumHeads  int
    NumLayers int
    FFNDim    int
    VocabSize int
    MaxSeqLen int
    Dropout   float64
}

Agent Framework (agents/framework/)

agent := agents.NewReActAgent(config, llmProvider)
result := agent.Run(ctx, "Your task here")

API Endpoints

Portal (Unified Platform)

• GET /api/dashboard/metrics - Training metrics
• GET /api/models - Model registry
• GET /api/marketplace/apps - Marketplace apps

Marketplace

• GET /api/apps - List apps
• GET /api/apps/search?q= - Search
• POST /api/apps - Register app
• GET /api/stats - Statistics

Model Registry

• GET /api/models - List models
• POST /api/models - Create model
• GET /api/models/:id/versions - List versions
• PUT /api/models/:id/versions/:v - Update stage

Design Principles

1. Zero Dependencies - Only Go standard library
2. Full Transparency - Complete algorithm implementations, no black boxes
3. Modular Design - Each module is independently importable
4. Production Ready - REST APIs, web UIs, persistence

Directory Structure

/workspace/
├── tensor/                 # Tensor data structures & math
├── nn/                     # Unified NN layers, optimizers, schedulers
├── transformer/            # Full Transformer with backpropagation
├── transformers/           # GPT-2 and BERT implementations
├── vector_db/              # HNSW vector database
├── asr/                    # Speech recognition (MFCC, LSTM, CTC)
├── finetune/               # Fine-tuning with LoRA support
├── agents/framework/       # AI agent framework
├── self_improving_agents/  # Recursive self-improvement
├── Advanced-Research/      # MoE, Diffusion, RL, GraphNN
├── rag_eval/               # RAG/LLM evaluation toolkit
├── dashboard/              # Training metrics dashboard
├── model_registry/         # Model versioning system
├── marketplace/            # MCP apps marketplace
├── portal/                 # Unified web platform
├── benchmarks/             # Performance benchmarks
├── opencode-lrs-agents-nbx/# Distributed training components
└── main.go                 # Root demo

Key Files

File	Purpose
tensor/tensor.go	Core tensor with gradient support
tensor/ops.go	Mathematical operations (matMul, softmax, etc.)
transformer/model.go	Complete encoder-decoder architecture
transformer/backward.go	Full backpropagation implementation
nn/layers.go	Linear, Embedding, LayerNorm layers
nn/optimizers.go	Adam, AdamW, SGD, Lion, RMSprop
vector_db/hnsw.go	HNSW ANN search index
asr/model.go	LSTM acoustic model
agents/framework/core/types.go	Agent, Tool, Message types
portal/main.go	Unified platform entry point

Benchmarks

Run the benchmark suite:

cd benchmarks
go test -bench=. ./...

Benchmarks cover tensor operations, transformer training, ASR pipeline, and vector DB performance.

License

MIT License - Free to use, modify, and distribute.

---

Built with pure Go - zero external dependencies