Executive Summary

The Problem

Large Language Models are being deployed in enterprise environments where factual accuracy is non-negotiable: healthcare diagnostics, financial compliance, legal contract review, and aviation safety. Yet every LLM hallucinates — fabricating facts, inventing citations, and misrepresenting data with confident-sounding prose.

Existing solutions focus on prompt injection defense, content moderation, or conversation guardrails. None of them answer the fundamental question: "Is what the model just said actually true according to my data?"

The Solution

TriniGard is a governance layer that sits between your data sources and your LLM provider. Before any LLM-generated text reaches your end user, TriniGard verifies factual claims against authoritative sources you configure: PostgreSQL databases, REST APIs, Elasticsearch clusters.

• Multi-source parallel verification with confidence scoring
• Per-use-case trust thresholds (0.65 to 0.98)
• Configurable fallback behavior: reject, flag for review, or proceed with warning
• Immutable write-ahead audit logging with HMAC signatures
• Vendor-independent — works with Claude, GPT, Llama, or any custom model
• Sub-300ms verification latency with Redis caching

Architecture

TriniGard is deployed as a lightweight API service alongside your existing infrastructure:

• Verification Engine: Queries configured data sources in parallel, computes confidence scores, applies threshold logic
• Data Source Adapters: PostgreSQL (complete), REST API (complete), Elasticsearch (planned). Abstract adapter pattern for any source.
• Trust Configuration: DB-backed per-use-case threshold management with hot-reload
• Audit Logger: WAL with HMAC signing + async PostgreSQL sync for SOC2/HIPAA compliance
• Rate Limiting: Token bucket algorithm, 60 RPM / 1000 RPH per tenant
• Result Caching: Redis-backed SHA256-keyed cache, 300s TTL

Target Markets

TriniGard is designed for regulated and high-stakes verticals:

• Healthcare: Verify medical claims against clinical guidelines and drug databases
• Financial Services: Validate compliance statements, risk assessments, regulatory filings
• Legal: Cross-reference contract clauses against case law and regulatory text
• Aviation: Safety-critical verification with zero-tolerance thresholds
• Public Sector: Government AI deployments requiring audit trails and explainability

Technical Specifications

• Stack: FastAPI (Python 3.11), asyncpg, Redis, PostgreSQL 15
• Performance: 29+ requests/second, sub-300ms p95 latency
• Testing: 47 E2E tests, 1500-request load test (100% pass rate)
• Deployment: Docker Compose, 4 services, <2 minute startup
• Auth: JWT with bcrypt, API key support
• Observability: Prometheus metrics, health checks (DB, Redis, audit, engine)
• Admin UI: React + Vite dashboard for use case management and audit review

Key Differentiators

• Not a guardrail: TriniGard verifies against real data, not patterns or prompts
• Not a wrapper: Vendor-independent architecture, no lock-in to any LLM provider
• Compliance-ready: Immutable audit trail designed for SOC2/HIPAA from day one
• Configurable risk: Set your own thresholds per use case, not a one-size-fits-all
• Production-tested: Load tested, E2E validated, Docker-native deployment