Claude India DPI
romichandra/claude-india-dpi
Gateway IA seguro para DPI indiano com validação neuro-simbólica
Ver no GitHubFramework enterprise-grade para escalar modelos de linguagem de forma segura e determinística na infraestrutura digital pública da Índia via MCP. Implementa validação com Pydantic, verificação neuro-simbólica com SymPy e avaliação multilíngue (inglês, hindi, tâmil) para eliminar alucinações de ferramentas e desvio semântico em ambientes de alto risco.
Instalação
npx convez-ai romichandra/claude-india-dpi1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Integração segura com registries MeitY API Setu
- Processamento determinístico de infraestrutura pública crítica
- Validação multilíngue em ambientes de alto risco regulatório
- Verificação matemática de operações STEM com SymPy
Veredito

Solução altamente especializada para ambientes de infraestrutura pública regulada na Índia. Complexa e pouco reutilizável fora do contexto de DPI indiano; requer profundo entendimento de segurança, MCP e validação neuro-simbólica.
README
A tradução para o português deste README ainda está na fila.
Claude-India-DPI: High-Stakes Public Infrastructure AI Gateway Framework
IndiaDPI (india-dpi) is an enterprise-grade, production-ready system architecture designed to safely, reliably, and deterministically scale frontier language models (Claude 4.5/Sonnet) across highly-regulated public digital infrastructure.
This framework mitigates non-deterministic LLM behaviors, runtime tool hallucinations, and cross-lingual semantic drift within high-stakes environments like MeitY API Setu registries and national public welfare distribution networks.
Core Architecture & Safety Layers
The framework implements a strict multi-tier safety pipeline to enforce isolation over LLM runtime boundaries:
1. Perimeter Type-Guards (mcp_server.py)
Exposes secure service abstractions to Claude using the Model Context Protocol (MCP). Tool parameters generated by the model are intercepted at the perimeter by strict Pydantic Validation Contracts before network processing, blocking syntax or parameter hallucinations at the edge.
2. Concurrent Cross-Lingual Evaluation Matrix (eval_engine.py)
An asynchronous batch processing engine that concurrently stresses system instructions across English, Hindi, and Tamil variants using asyncio.gather and ephemeral prompt caching. An isolated, schema-enforced model auditor programmatically scores semantic drift and logic omissions.
3. Neuro-Symbolic Verification (neuro_symbolic.py)
A hybrid processing paradigm for zero-tolerance math/STEM applications. It treats the neural network (Claude 4.5 Haiku) as an untrusted proposal generator for unstructured text parsing, while using a deterministic SymPy symbolic execution engine to mathematically verify algebraic equivalence before allowing execution states to update.
4. API Setu Production Client (apisetu_client.py)
A decoupled, asynchronous network transport manager built to connect directly to India's national data-sharing platform (MeitY API Setu). It wraps downstream API calls inside thread-safe connection pools with rigid resource limits to eliminate socket exhaustion risks under high concurrent load profiles.
System Requirements & Tooling Stack
IndiaDPI is engineered to eliminate local machine state dependencies. It runs on a fully isolated workspace driven by the modern Rust-backed Python manager uv.
- Runtime: Python 3.10+
- Dependency Engine:
uv(replaces traditional pip/venv setups with hard-linked caching) - Core Libraries:
mcp,pydantic-settings,httpx,sympy,loguru,pytest-asyncio
Installation & Quick Start
1. Clone and Sync the Environment
Clone the repository and compile the completely isolated dependency tree instantly using uv:
git clone https://github.com/romichandra/claude-india-dpi.git
cd india-dpi
uv sync
2. Configure Your Environment Variables
Create a local .env configuration file in the project root folder:
cp .env.example .env
Open .env in your editor and supply your validated authentication credentials:
ANTHROPIC_API_KEY="sk-ant-your-key-here"
APISETU_CLIENT_ID="your_apisetu_client_id"
APISETU_API_KEY="your_apisetu_secret_key"
APISETU_USE_SANDBOX=true
# Dynamic Execution Routing Matrix
MCP_RUNTIME_MODEL="claude-haiku-4-5-20251001"
EVAL_TARGET_MODEL="claude-sonnet-4-6"
EVAL_GRADER_MODEL="claude-sonnet-4-6"
3. Launch the Local Protocol Inspector
Verify the MCP data gateway integration visually using the global protocol inspector:
npx -y @modelcontextprotocol/inspector uv run src/india_dpi/mcp_server.py
Production Telemetry Design
Telemetry Operational Note:
IndiaDPI enforces an absolute print isolation rule. Standard stdout streams are preserved exclusively for raw JSON-RPC 2.0 protocol packets. All internal application monitoring, performance counters, and framework tracebacks are intercepted by loguru and explicitly routed to sys.stderr to avoid communication channel corruption.
Directory Architecture
src/india_dpi/
├── __init__.py
├── config.py # Centralized settings engine parsing environment variable models
├── apisetu_client.py # Asynchronous connection pool interface for MeitY API Setu
├── mcp_server.py # Schema-enforced Model Context Protocol gateway
├── eval_engine.py # Asynchronous parallel cross-lingual evaluation harness
└── neuro_symbolic.py # Hybrid neural-parsing/symbolic-math validation engine