Fable 5.2 notices when your gateway edits the conversation
Fable 5.2 preserved thinking ties replayed reasoning to the exact history, the model and the account. What that asks of an AI gateway, and what it costs.
Writing
Field notes on AI platform architecture
Design guides from real platforms: what to build, in what order, and what it costs you.
Don't give your AI gateway the keys
AI gateway security after three exploited LiteLLM flaws: where provider keys should live, which features to switch off, and five tests that it fails closed.
What the 2026-07-28 MCP specification update changes for agents that write
The 2026-07-28 MCP specification update, read against six write-access requirements: what it now says about approvals and retries, and what stays yours.
What to require from an MCP server before an agent can write to your systems
MCP server security for agent write access: six requirements to put to any gateway or action server, and how to test each one in a demo.
Building an enterprise AI gateway on Amazon Bedrock
Design an AI gateway on Amazon Bedrock rule by rule: EU residency, one path to models, Guardrails, routing, quotas and logs, each with a test to run.
Enterprise AI platform architecture: a reference guide to the five layers
Enterprise AI platform architecture, layer by layer: what each layer decides, where to enforce it, how to test it, what it costs and when to build it.
Human-in-the-loop agent actions over MCP: idempotency, audit and replayable decisions
How to build human-in-the-loop agent actions over MCP: approvals bound to exact parameters, idempotency keys, preconditions and a decision log, each tested.
LLM evaluation in CI: golden sets, DeepEval and Langfuse
LLM evaluation in CI, decision by decision: golden set cases, DeepEval metrics, a baseline with measured noise, calibrated judges and Langfuse traces.
RAG that cites its sources or declines: pgvector, hybrid search and Row-Level Security
RAG with citations on PostgreSQL: where each guarantee is enforced, the Row-Level Security policy, recall under a tenant filter and how to test each.
Last updated:
Got an AI project stuck between demo and production?
Tell me what you are building. After a first conversation you will know whether I can help and what I would do first.
+33 6 02 73 49 22 me@osmanrami.fr