corvus-ai 0.3.78

Platform-agnostic, extensible AI-powered ML Development Assistant

Corvus - ML Development Assistant

Platform-agnostic, extensible AI-powered ML Development Assistant

Overview

Corvus is built on three pillars:

1. Universal Skills Framework - 24 ML skill functions across 24 modules for data engineering, model training, deployment, observability, and ML advisory
2. MCP Integration - 6 MCP server integrations for GitHub, PostgreSQL, Filesystem, MLflow, Docker, and Cloud Storage
3. Knowledge Base - Hybrid retrieval (BM25 + Vector + RRF), GraphRAG with Neo4j, persistent memory with Mem0
4. AI Safety - Defense-in-depth security with prompt injection defense, DLP, grounding validation, and quality monitoring

Features

Skills Library (24 Skills)

Category	Skills
Data Engineering	data-profiler, feature-recommender, pipeline-generator, schema-designer, validation-generator
Model Training	model-selector, architecture-recommender, pytorch-codegen, experiment-tracker, distributed-config, hyperparameter-optimizer
Deployment	endpoint-generator, docker-generator, terraform-generator, k8s-manifest-generator, cicd-generator, quantization-exporter
Observability	monitoring-setup, drift-detector, dashboard-creator
ML Advisory	ml-spec-generator, experiment-advisor
Model Hub	huggingface-skills, github-ml-skills

MCP Integrations (6 Servers)

Server	Capabilities
GitHub	Repository access, PRs, issues, commits
PostgreSQL	Database queries, schema management
Filesystem	File operations, directory traversal
MLflow	Experiment tracking, model registry
Docker	Container management, image operations
S3/GCS	Cloud storage operations

Knowledge Base

Component	Description
Qdrant Vector Store	Semantic search with voyage-code-3 embeddings
Hybrid Retrieval	BM25 + Vector search with RRF fusion
Cross-Encoder Reranking	Precision improvement for retrieval
Neo4j GraphRAG	Knowledge graph with entity extraction
Mem0 Memory	Persistent memory across sessions

AI Safety & Guardrails

Component	Purpose
Prompt Guard	Injection detection and input validation
Tool Guard	MCP security and sandboxing
Memory Guard	Memory injection prevention
Grounding Validator	Hallucination prevention with citations
DLP Scanner	PII/credential detection and redaction
Quality Monitor	Output quality and drift detection

Quick Start

# Clone
git clone https://github.com/CloudlyIO/corvus.git
cd corvus

# Install with all dependencies
pip install -e ".[all]"

# Install pre-commit hooks
pre-commit install --hook-type pre-commit --hook-type commit-msg

# Run tests
make test

Usage Examples

Using Skills

import asyncio
from corvus.skills.library.model_selector import select_model
from corvus.skills.library.pytorch_codegen import generate_pytorch_code

async def main():
    # Select a model for your task
    result = await select_model(
        task_type="classification",
        data_characteristics={
            "num_samples": 10000,
            "num_features": 50,
            "num_classes": 3,
        },
    )
    print(f"Recommended: {result['recommended_model']}")

    # Generate PyTorch training code
    code = await generate_pytorch_code(
        model_type="mlp_classifier",
        input_features=50,
        output_classes=3,
        hidden_layers=[128, 64],
    )
    print(code["model_code"])

asyncio.run(main())

Using Safety Features

from corvus.safety import (
    validate_input,
    create_dlp_scanner,
    assess_quality,
)

# Validate user input for injection attempts
result = validate_input("Help me build a classifier")
if result.is_safe:
    print("Input is safe")

# Scan for sensitive data
scanner = create_dlp_scanner()
dlp_result = scanner.scan_content("Contact: [e-mail]")
if dlp_result.pii_count > 0:
    print("PII detected!")

# Assess response quality
quality = assess_quality(
    "Here's a detailed explanation with examples...",
    prompt="Explain gradient descent"
)
print(f"Quality score: {quality.overall_score:.2f}")

Running the E2E Demo

# Run the complete ML workflow demo
python3 demos/e2e/ml_workflow_demo.py

Development

# Testing (TDD)
make test           # Unit tests
make test-all       # All tests
make test-cov       # With coverage

# Evaluation (EDD)
make benchmark      # Run benchmarks

# Quality
make lint           # Ruff linting
make format         # Black formatting
make quality        # Lint + typecheck

Project Structure

corvus/
├── src/corvus/              # Main package
│   ├── core/              # Config, exceptions
│   ├── skills/            # Universal skills framework (24 skills)
│   ├── llm/               # LLM gateway (LiteLLM)
│   ├── agents/            # Agent orchestration (LangGraph)
│   ├── knowledge/         # Knowledge base (RAG/GraphRAG/Memory)
│   ├── mcp/               # MCP integrations (6 servers)
│   └── safety/            # AI safety & guardrails (6 components)
├── tests/                 # Test suite
│   ├── unit/              # 10,168+ tests
│   ├── integration/       # integration tests
│   └── e2e/               # End-to-end tests
├── benchmarks/            # Performance benchmarks
├── demos/                 # Executable examples
│   ├── skills/            # Individual skill demos
│   ├── knowledge/         # Knowledge base demos
│   ├── mcp/               # MCP integration demos
│   ├── safety/            # Safety feature demos
│   └── e2e/               # End-to-end workflow demos
└── docs/                  # Documentation

Documentation

Document	Description
Documentation Index	All guides, API reference, architecture docs
Status	Current project status
Architecture	Technical architecture
Contributing	Developer workflow
Changelog	Version history

Configuration

# Core
CORVUS_DEBUG=true|false
CORVUS_LOG_LEVEL=INFO|DEBUG|WARNING|ERROR

# LLM
CORVUS_LLM_DEFAULT_PROVIDER=anthropic|openai
CORVUS_LLM_DEFAULT_MODEL=claude-sonnet-4-20250514

# Knowledge Base
CORVUS_KNOWLEDGE_VECTOR_STORE=qdrant
CORVUS_KNOWLEDGE_EMBEDDING_MODEL=voyage-code-3

Technology Stack

Layer	Technology
Agent Orchestration	LangGraph
LLM Gateway	LiteLLM
Primary LLM	Claude Sonnet 4 / Opus 4.5
Vector Database	Qdrant
Knowledge Graph	Neo4j
Code Embeddings	voyage-code-3
Text Embeddings	e5-large-v2
Memory System	Mem0
Tool Integration	MCP

License

MIT