Skip to main content

​
The 13-step request lifecycle

  1. Parse request body.
  2. Per-IP global rate limit (pre-auth).
  3. Authenticate (API key or JWT).
  4. Per-tenant rate limit.
  5. PII scan.
  6. Prompt-injection scan (parallel with 5).
  7. Policy resolution → ALLOW / REDACT / BLOCK / ESCALATE.
  8. Enforce.
  9. Forward to LLM (Groq, Gemini, Mistral, Ollama).
  10. Response compliance check.
  11. Write HMAC-signed, hash-chained audit row.
  12. Nightly Merkle anchor (S3 Object Lock).
  13. Deliver response with X-AGCMS-Interaction-ID.

​
Multi-tenancy

Postgres row-level security partitions every table by tenant_id. The gateway sets current_setting('app.current_tenant') on each connection. There is no shared global query path that bypasses RLS — including the audit log.

​
The audit chain

Each row’s HMAC-SHA256 signature is computed over (payload || previous_log_hash), and the row’s previous_log_hash equals the prior row’s signature. Per-tenant sequence numbers are assigned via a Postgres advisory lock so rows are contiguous within a tenant. Every night a per-tenant Merkle tree is built over the day’s signatures; the root is signed with a separate anchor key and pushed to an S3 Object-Lock bucket in Compliance mode (default 7-year retention).

​
Why this design wins compliance reviews

  1. Truncation, reordering, or row substitution breaks the chain.
  2. Substituting a Merkle root requires bypassing S3 Object Lock.
  3. The bundle exporter ships a self-contained verify.py — auditors never need AGCMS credentials to validate integrity.