Code-knowledge system for an enterprise fintech codebase

Separate facts from interpretation

A static-analysis worker built on headless IntelliJ provides the exact structure of the code. The AI never guesses this part. Its job is to explain the structure: business meaning, decisions, rules, and the SQL that the code really builds.

• Classes, methods, call graphs and real entry points
• Decision trees: every branch and its arms
• SQL, messaging and HTTP call sites

A team of agents with fresh context

A lead agent plans the work and hands slices of coverage to worker agents. Each worker starts with a fresh context, so long sessions do not degrade. Agents use read-only tools: the code-analysis worker, reference data, version control, the issue tracker and the wiki.

Tools check the agent, the agent checks the tools

• Coverage: a deterministic list names everything that must be explained, including flows, methods, decision sites, SQL sites and coded values. Training is not finished until each item is explained or acknowledged with a reason.
• Validation: every write is checked against code facts. A wrong id, evidence outside the method, or a constant name in place of SQL is refused.
• SQL: static detection is treated as a lower bound. The agent rebuilds the SQL from code fragments, and an independent scan catches what both missed.

A versioned knowledge graph

Knowledge is stored in Neo4j and versioned by generation. Each publish switches the view atomically, and a failed publish rolls back. Shared resources such as tables, columns, topics and endpoints have one identity, so services link to each other in any training order.

Qdrant holds embeddings, which are used only to find starting points for vague questions.

Retrieval as graph planning

The retrieval server is a planner over the graph. It parses a question into an intent and entities, resolves them to graph nodes, and follows the relevant edges: rules to decisions to SQL to tables, or a trace across services. It returns a compact answer with citations. Results are paged and never truncated.

It runs as an MCP server, so engineers use it from Claude Code, VS Code or Claude Desktop.

Built to be operated

• Incremental training: after the first full run, only changed methods are learned again on each new commit.
• Resumable stages: clone, setup, skeleton, learn and publish are recorded, so a stopped run resumes where it stopped.
• Parallel trainers share the stores safely through leases and conditional claims.
• Runs as 3 containers with Podman and Docker on WSL.
• Tested before spending on real agent runs: scripted-agent rehearsals, chaos tests (kill -9, SIGTERM, budget stops) and a scale rehearsal with parallel trainers.