Show HN: Inspect Element for LLM Apps (github.com)

🤖 AI Summary
SATI is a lightweight protocol and reference implementation that treats LLM interactions as first-class HTTP/REST resources: every model call is an immutable, timestamped, verifiable record so conversations can be reconstructed statelessly from logs rather than held in session stores. The demo shows local (Ollama) and cloud (OpenAI/Anthropic) setups with identical UIs, and exposes standard web toolchains—Prometheus, OpenTelemetry, API gateways, tracing, and load balancing—against LLM traffic. Integrity is enforced via blockchain-style chain hashing, and the system surfaces structured reasoning traces so you can inspect not just outputs but the internal steps that produced them. For engineers and platform teams this matters because it makes LLM middleware vendor-neutral, observable, auditable, and compliance-ready out of the box: swap providers without refactoring, replay any turn with full context, and present tamper-evident logs suitable for regulated domains. The design is intentionally modular—pluggable RAG/vector stores, stronger hashing/HSMs, policy layers (rate limits, filters) and standard auth (OAuth, mTLS) are all compatible because every interaction is an HTTP request. MIT-licensed and submitted to ISED’s AI sprint, SATI demonstrates a practical, production-friendly path to monitorable, auditable LLM systems using existing web infrastructure.
Loading comments...
loading comments...