Ingestion
The Document Loaded. The Meaning Didn't.
Every ingestion check passed. The answers still mixed up two vendors' instructions.
Post-mortem analysis for AI architects
Reason forOutage.
Real failures.
Full traces.
Architecture decisions.
Post-mortems for agentic AI systems, written by the architect who built one. I run an enterprise agentic RAG platform for a global enterprise. When it breaks, I trace what happened, find the cause, and write up the fix. You get the lesson without the outage.
Read the post-mortems ↗RFO / Anatomy of an agentic system
The model is one part.
Evidence, tool use, and the boundaries around action. Every post-mortem starts with the traces.
One real failure,
followed all the way down.
Each post-mortem takes a single production incident through the whole agentic RAG stack, dead ends included. If I cannot reconstruct what the system did from its traces, I cannot explain the failure, so the first fix is often better observability. Technical enough to reuse, plain enough to forward to your manager.
- 01 / Symptom
What did they see?
What the user or operator saw, in their words.
- 02 / Timeline
What did the system do?
Reconstructed from the traces.
- 03 / Root cause
Where did it break?
Ingestion, chunking, retrieval, reranking, generation or tool use.
- 04 / Blind spot
Why was it missed?
Why monitoring missed it and what I measure now.
- 05 / Fix
What changed?
What I changed, what I rejected, and how I proved it worked.
- 06 / Decision
What should you decide?
The architecture choice, so you can make your own.
↗
Agents
Your Agent Answered. Did It Do the Job?
A fluent answer, a tidy References section, and four links that did not exist.
Architecture
Five Levels of RAG, and the Failure That Ends Each One.
The demo was the easy part. Five failure modes that changed how I built the system.
Written by the person who built itArchitect and builder,
Architect and builder,
not a case study.
I designed and built an agentic RAG orchestration layer for a global enterprise, and I wrote the code. The platform is fully traced, with golden evals behind every change.
I came up through network engineering, so I care most about what happens after deployment.
Retrieval, ingestion, agents.
The design decisions and the boundaries between components.
02 / Failure analysisFollow the failure.
What broke, what the traces showed, and what reduced the risk.
03 / ObservabilityKnow what it did.
Tracing, golden evals, and review gates around AI systems.
Reason for OutagePost-mortems for
Post-mortems for
AI architects.
A post-mortem is a write-up of what went wrong and what changed. New ones land in your inbox, written to be read in five minutes.
Subscribe to RFO ↗