LLMs are stateless. REST is stateless. SATI is a protocol that makes LLM conversations work like REST resources.

Treat LLM interactions like HTTP requests—every call is logged, timestamped, and verifiable. This unlocks standard web tooling: Prometheus monitoring, OpenTelemetry tracing, API gateways, and the DevOps stack you already use.

Chat interface on the left. Full observability on the right.

• Stateless reconstruction - no session store, conversations rebuilt from immutable records
• Tamper-proof history - blockchain-style chain hashing for integrity
• Structured reasoning - see what the AI was thinking, not just what it said

Observe, audit, and verify every LLM interaction using standard HTTP/REST patterns.

Local (Ollama):

# Node.js
ollama pull mistral
cd nodejs && npm install && node server.js

# .NET
cd dotnet && dotnet run

Cloud APIs (OpenAI/Anthropic): Edit config.json with your provider and API key, then run.

Both run on http://localhost:3000 with identical UI.

Core Benefits:

• Provider independence - Switch from OpenAI to Anthropic to local models by without refactoring
• Observability as infrastructure - Monitor LLMs like you monitor REST APIs
• Compliance-ready audit trails - Cryptographic chain proves what the AI said (legal/medical/financial)
• Stateless by design - Replay any turn with full context, no session state to manage.
• Prompt injection resilience - Instructions regenerated per turn, not hijacked

Who This Is For:

Developers prototyping AI features - Ship in hours using local models, migrate to cloud APIs later without changing middleware. Drop in your files, modify instructions and run it.

Platform teams - Vendor-neutral telemetry, request/response tracing, and policy enforcement at the HTTP layer.

Regulated industries - Tamper-evident audit logs work identically whether you're using Azure, OpenAI or self-hosted models.

Full technical deep-dive: pattern.md

SATI demonstrates that LLM middleware can be done simply with HTTP.

This is intentionally minimal. Each abstraction is intentionally modular so it can be improved.

• Swap in production-grade RAG (vector databases, reranking)
• Upgrade hashing (HMAC, HSM integration)
• Add policy layers (rate limiting, content filtering)
• Implement auth (OAuth, API keys, mTLS)

Every interaction is HTTP, so standard web patterns just work: OpenTelemetry tracing, API gateways, load balancers, CDN caching. Fork it. Build with it. Make it yours.

MIT

Submitted to the ISED Canada AI National Sprint as a demonstration of HTTP-native patterns for observable, auditable LLM systems.