Frontier Tech / AI — RAG Documentation & LLM Log-Triage

250+

Pages of Documentation Autonomously Restructured

98%

Structural Accuracy on RAG-Generated Docs

100,000+

Monthly Failure Events Auto-Clustered

The Problem

Platform documentation for the crypto reporting system was manually authored and rapidly falling out of date. Simultaneously, the on-call support team faced 100,000+ monthly distributed failure events with no automated way to cluster or triage them, severely slowing incident resolution during peak failure volumes.

The Architecture

Working across the engineering organization, we engineered a Claude- and OpenAI-powered RAG pipeline that autonomously structures and publishes platform documentation directly to Confluence. Concurrently, we developed an LLM-driven log-triage agent (Claude, OpenAI, Codex) integrated natively with Datadog telemetry to autonomously cluster failure events for the on-call rotation.

The Constraints

Zero existing internal patterns for frontier-model tooling. Outputs had to be exceptionally reliable—hallucinations were unacceptable—to eliminate manual authoring and actively reduce, rather than add to, on-call cognitive load.

The Outcome

Eliminated thousands of manual documentation authoring hours for 50+ engineers, accelerating stakeholder approvals for continued platform investment. Drastically cut manual log analysis and accelerated MTTR (Mean-Time-To-Resolution) across the on-call rotation without escalating to core engineering.