An engineering organization
Institutional memory for humans and coding agents
role: Designed and built the LLM client, the ingestion and review agents, and the retrieval layer.
// problem
Architectural knowledge lived in people's heads and in merged PRs. Coding agents started every session without context and kept repeating the same mistakes.
// approach
- A provider-agnostic LLM client with a bounded tool-use loop.
- Ingestion agents run on merged PRs, Jira and Confluence. A relevance gate decides what becomes institutional knowledge, committed as structured, signed markdown.
- A retrieval layer of 200+ knowledge files (index, quick reference, flow traces), served locally to Claude Code, Cursor, Copilot and Codex, and remotely through a containerized MCP server.
- A code-review agent that grounds its reviews in the same knowledge tools.
// architecture
Merged PRs
Jira
Confluence
Ingestion agent
Relevance gate
Knowledge repo
signed markdown
MCP server
Local files
Coding agents
Review agent
// guardrails and production concerns
- The tool-use loop is bounded.
- A relevance gate filters what gets committed, so the knowledge base stays useful instead of growing noisy.
- Knowledge is plain markdown in a repository: reviewable, versioned and signed.
// outcome
Agents and engineers share one source of truth that stays current. AI-assisted review became a standing control, and it surfaced a supply-chain security issue.
// stack
- AI agents
- Tool use
- MCP
- Retrieval
- Claude Code
- Cursor
- Copilot
- TypeScript