An intelligent, evidence-grounded brand guardrail system that evaluates marketing copy, press releases, social drafts, and AI-generated content against approved brand guidelines and verified facts before publication.
BrandShield AI is a pre-publication brand governance engine. Unlike post-publication monitoring tools that scan Twitter or websites after a crisis occurs, BrandShield operates at drafting time inside the copywriter’s, marketer’s, and AI agent’s workflow to ensure that nothing non-compliant, unapproved, or legally hazardous ever goes live.
Generic LLMs hallucinate phantom rules. In BrandShield, the Verifier Gatekeeper Agent automatically drops any flag that cannot cite an exact clause in the active Brand Pack (e.g. Brand Glossary §1.1 or Fact f_pricing_001). If there's no citation, there is no flag. Period.
When fixing violations, generative AI often changes user intent or introduces new prohibited promises (like turning "the best" into "100% guaranteed returns"). The Intent Guard Engine analyzes semantic coherence and negations to ensure every suggestion is safe before presenting it to the user.
Native Model Context Protocol (MCP) server integration. External AI systems (Claude Desktop, Cursor, ChatGPT plugins, marketing scripts) can call check_content and lookup_fact to govern content drafts automatically.
With generative AI, companies produce 100x more content across dozens of agencies, marketing teams, and regional offices. But brand guidelines remain static 80-page PDF documents that humans ignore and LLMs hallucinate over.
In fintech, claiming "guaranteed returns", "risk-free deposit", or "100% safe money" violates RBI, SEBI, and FTC regulations. A single rogue LinkedIn post can trigger regulatory audits, license freezes, and severe statutory fines.
A marketing copywriter uses a ChatGPT prompt that quotes last quarter's pricing (e.g. ₹4,999/mo instead of ₹3,999/mo) or claims an SLA (99.9%) that was superseded. Customers sign contracts based on false promises, forcing legal concessions.
Writing "Novapay" instead of "NovaPay", using banned hyperbolic slang like "insane savings that will blow your mind!!", or disrespecting competitors erodes the enterprise trust that took years to cultivate.
Social images and banners generated with insufficient contrast ratios fail WCAG 2.2 AA standards (< 4.5:1), exposing brands to accessibility lawsuits and alienating visually impaired users.
| Evaluation Dimension | Manual Human Review | Generic LLM Prompt | BrandShield AI |
|---|---|---|---|
| Detection Latency | 2-48 hours backlog | 2,500ms - 5,000ms | 20.4ms (Hybrid) |
| Citation Grounding | Selective / inconsistent | Hallucinates non-existent rules | 100% verified clause citation |
| Fix Suggestion Safety | Prone to typos | Introduces new hallucinated promises | Intent Guard checked & verified |
| Audit Trail Integrity | Slack threads, email trails | Ephemeral chat logs | SHA-256 tamper-proof ledger |
| External Agent Access | None | Manual copy-paste | Brand Guard MCP (JSON-RPC) |
BrandShield enforces strict Enterprise Role-Based Access Control (RBAC). In the top navbar of the app, click any persona pill to instantaneously transform the user experience, active permissions, and visible sidebar tabs.
Title: Senior Marketing Copywriter · Desk: Marketing Studio
Q3 Enterprise Launch Announcement — LinkedIn from the dropdown or paste fresh marketing copy.
🔒 LockedThe application is pre-seeded with rich enterprise content items demonstrating how each of the specialized agents catches, verifies, and fixes real-world brand risks.
Content ID: cnt_demo_fintech_launch · Channel: LinkedIn · Target: B2B CFOs
| Violation String | Detecting Agent | Severity | Exact Guideline Citation | Intent-Safe AI Fix |
|---|---|---|---|---|
| "Novapay" | Terminology Agent | CRITICAL | Brand Glossary §1.1 | "NovaPay" (Proper camel-case) |
| "the best payment platform in India" | Tone & Voice Agent | WARNING | Brand Voice Guide p.8, §2.1 | "a leading payment solution for Indian enterprises" |
| "₹4,999/mo" | Claims & Facts Agent | CRITICAL | Fact f_pricing_001 (Pricing Sheet 2026 v2.4) | "₹3,999/mo" (Current verified active price) |
| "guaranteed returns" | Messaging & Compliance Agent | CRITICAL | Regulatory Glossary §4.2 (RBI Guardrail) | "transparent yield management across corporate escrow" |
| "insane savings that will blow your mind!!" | Tone & Voice Agent | WARNING | Brand Voice Guide p.9, §2.3 | "predictable and competitive transaction fee structures" |
When an unsafe AI fix like "a platform with 100% guaranteed returns" was proposed for the original text "the best", the Intent Guard engine intercepted and flagged it immediately because it introduced a banned financial promise. Only verified, safe suggestions reach the user's screen!
Model Context Protocol (JSON-RPC 2.0) · Tool: check_content
{
"mcpServers": {
"brandshield": {
"command": "python",
"args": ["-m", "apps.api.mcp.brand_guard_mcp"],
"env": {
"PYTHONPATH": "d:\\nova"
}
}
}
}
check_content):{
"governance_score": 52,
"risk_tier": "HIGH",
"citations_verified": true,
"findings_count": 5,
"gatekeeper_status": "PASSED_GROUNDING",
"findings": [
{
"rule_id": "r_term_001",
"citation": "Brand Glossary §1.1",
"suggestion": "NovaPay"
}
]
}
BrandShield replaces fragile monolithic prompt chains with a deterministic-first, LangGraph-orchestrated multi-agent pipeline. Each agent has an isolated responsibility and strict Pydantic contract.
Extracts text spans, sanitizes PII, calculates sentence boundaries, and normalizes unicode characters.
Adjusts risk weights based on channel (LinkedIn, Press Release vs internal memo) and target audience (CFOs, retail).
Exact CamelCase and banned term detection using regex and phrase matcher trees (e.g. Novapay → NovaPay).
Calculates relative luminance contrast ratios (WCAG 2.2 AA target 4.5:1) and Flesch-Kincaid grade levels.
Validates hex colors against approved palette and checks logo clear-space buffer ratios.
Catches unsubstantiated superlatives, hyperbole, and aggressive slang while enforcing measured confidence.
Cross-references numerical, SLA, and pricing claims against verified enterprise facts in the Brand Pack.
Flags deprecated product lines and guarantees messaging reinforces current strategic positioning.
THE "No citation, no flag" barrier. Kills uncited flags and hallucinated rules with 100% mathematical precision.
Synthesizes 2-3 tracked-change suggestions with editorial rationale and character offset tracking.
Blocks fix suggestions that introduce new ungrounded promises or invert negative constraints.
Computes 0-100 Governance Score and routes High Risk items exclusively to Legal Reviewers.
13a. Guideline Compiler: Transforms raw text and PDF documentation into structured JSON rules with positive/negative examples.
13b. Guideline Clarifier: Analyzes telemetry from reviewer overrides to identify ambiguous rules and drafts clarified wording for Admin approval.
Our evaluation suite tests both positive violation catches and negative control precision across an 80-item Gold Benchmark (50 violations, 30 clean controls).
python -u tests/master_test_suite.py
One-click links to test every part of the running solution.
Test editor canvas, dynamic text replacement, and inline citations.
http://localhost:3000Test Manager triage and Legal Reviewer high-risk override workflows.
http://localhost:3000/queueExplore 17 rules, 14 facts, Guideline Compiler & Clarifier.
http://localhost:3000/brand-packInteractive REST endpoints, SSE streaming routes, and schemas.
http://localhost:8000/docs