Deep-tech for responsible AI.
We build developer tools, publish open research, and shape AI governance — with a first-class focus on India's regulatory landscape. LLM safety, compliance (GDPR, EU AI Act, India DPDP, HIPAA), red-teaming, and safe regeneration.

Website · Docs · PyPI · npm · MCP Server · Discussions

_{If our work helps you ship safer AI, a star on rail-score-sdk keeps the roadmap funded.}

A deep-tech AI startup making artificial intelligence safer and more accountable. We operate across three fronts:

Built in India, built for everywhere. Reach us at responsibleailabs.ai.

RAIL Score is an API-first platform that evaluates AI-generated content across 8 independent dimensions on a 0-10 scale:

Fairness · Safety · Reliability · Transparency · Privacy · Accountability · Inclusivity · User Impact

Each evaluation returns per-dimension scores, confidence levels, issue detection, and actionable improvement suggestions. Not a single opaque number — a full diagnostic.

Built for teams shipping production LLM applications who need measurable, auditable safety, guardrails, and regulatory compliance inside their existing stack: drop-in provider wrappers, middleware, policy thresholds, multi-turn sessions, agent tool-call evaluation, prompt-injection detection, and single-call compliance across GDPR, HIPAA, EU AI Act, CCPA, India DPDP, and India AI Governance.

flowchart LR
    App["Your App / SDK"] --> RAIL["RAIL Score API"]
    MCP["MCP Client<br/>Claude · Cursor · Copilot"] --> RAIL
    RAIL --> Eval["Evaluate<br/>8 dimensions"]
    RAIL --> Comp["Compliance<br/>GDPR · HIPAA · EU AI Act<br/>DPDP · CCPA · India AI Gov"]
    RAIL --> RT["Red-team<br/>prompt injection · agent eval"]
    Eval --> Gate{thresholds met?}
    Comp --> Gate
    RT --> Gate
    Gate -- yes --> Pass["✓ Ship to user"]
    Gate -- no --> Regen["Safe-regenerate<br/>evaluate → improve → retry"]
    Regen --> Eval
    classDef brand fill:#75BFAF,stroke:#2D6B5A,color:#0A2A22;
    class RAIL brand;

Evaluation — Score any AI output with basic (~1 credit) or deep (~3 credits) analysis, with custom dimension weights for healthcare, finance, legal, and general contexts.

Safe Regeneration — Evaluate → improve → regenerate loop that runs until your thresholds are met. Server-side or client-side with your own LLM.

Compliance Testing — 63 requirements across 6 frameworks: GDPR, HIPAA, EU AI Act, CCPA, India DPDP Act, India AI Governance. Single API call.

Agent Evaluation — Tool call evaluation (ALLOW/FLAG/BLOCK), tool result scanning with PII redaction, and prompt injection detection across 5 attack patterns.

MCP Server: A hosted Model Context Protocol endpoint that adds the full RAIL safety layer to any MCP client (Claude, ChatGPT, Cursor, Copilot, Replit Agent, LangGraph, CrewAI) through a single URL, with no SDK integration. Exposes 8-dimension evaluation, agent tool-call gating, PII scanning, and India DPDP compliance.

Policy Engine & Middleware — RAILMiddleware wraps any generate function. PolicyEngine enforces per-dimension thresholds. RAILSession tracks multi-turn risk.

Configuration & Monitoring — Read the application's governance policy, plan capabilities, and dimension settings at runtime (get_config, get_capabilities, get_dimensions) for startup checks and observability.

LLM Observability — Native integration with Langfuse scores, LiteLLM guardrails, OpenAI, Anthropic, and Gemini.

Package	Registry	Install
Python SDK		pip install rail-score-sdk
JavaScript/TypeScript SDK		npm install @responsible-ai-labs/rail-score
Drupal Module		composer require drupal/rail_score
MCP Server		https://mcp.responsibleailabs.ai/mcp

pip install rail-score-sdk[openai]       # GPT-4o, GPT-4o-mini
pip install rail-score-sdk[anthropic]    # Claude 3.5, Claude 3
pip install rail-score-sdk[google]       # Gemini 1.5 Pro/Flash
pip install rail-score-sdk[litellm]      # LiteLLM guardrail support
pip install rail-score-sdk[integrations] # All providers

from rail_score_sdk import RAILClient

client = RAILClient(api_key="your-api-key")

result = client.evaluate(
    content="Your AI-generated text here",
    mode="deep",
    domain="healthcare"
)

for dim in result.dimensions:
    print(f"{dim.name}: {dim.score}/10 (confidence: {dim.confidence})")

import { RAILClient } from '@responsible-ai-labs/rail-score';

const client = new RAILClient({ apiKey: 'your-api-key' });

const result = await client.evaluate({
  content: 'Your AI-generated text here',
  mode: 'deep',
  domain: 'healthcare'
});

result.dimensions.forEach(dim => {
  console.log(`${dim.name}: ${dim.score}/10`);
});

from rail_score_sdk.wrappers import RAILOpenAI

client = RAILOpenAI(api_key="your-api-key", openai_api_key="sk-...")
# Every response is automatically scored
response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Summarize this report"}]
)
print(response.rail_score)  # Full 8-dimension evaluation attached

RAIL Score tests 63 requirements across 6 regulatory frameworks in a single API call:

RAIL Score is the only evaluation platform where India DPDP Act and India AI Governance are native, first-class compliance frameworks alongside GDPR, HIPAA, EU AI Act, and CCPA.

• India-first compliance, globally competitive — DPDP Act and India AI Governance ship as native frameworks, not bolted-on add-ons
• Full diagnostic, not a single score — eight independent dimensions, each with confidence, issues, and suggestions
• Closed-loop, not read-only — safe regeneration, policy enforcement, and session-level risk tracking, not just measurement
• Fits where developers already work — drop-in wrappers for OpenAI, Anthropic, Gemini, LiteLLM, and Langfuse; middleware for any generate function
• Grounded in research — peer-reviewable methodology on arXiv (2505.00204) and three public benchmarks on HuggingFace, including the RAIL Guard Benchmark for LLM and agent safety

• Questions, feature requests, how-to threads — GitHub Discussions
• Bugs and reproducible issues — open an issue
• Contributing — see CONTRIBUTING.md; issues labelled good first issue are the easiest starting points
• Security disclosures — see each repo's SECURITY.md

_{RAIL  ·  Responsible AI Labs Pvt. Ltd.  ·  Built in India, built for everywhere.}