Galaxium Onboarding Copilot

Every time a developer joins a new project, they lose real time just figuring out how the codebase works, tracing logic across files, hitting undocumented quirks the hard way, and interrupting teammates with basic questions, because documentation is usually missing or stale. This is slow, inconsistent, and genuinely frustrating for both the new hire and the team supporting them. I built an onboarding pipeline in IBM Bob that fixes this by turning any codebase into a ready-made onboarding kit automatically. Using IBM's Galaxium Travels demo app (an interplanetary flight booking system with a React/TypeScript frontend and a dual REST+MCP Python backend) as my test project, I built a custom, reusable Bob mode that generates: - Persistent project context via Bob's /init command - Architecture and sequence diagrams (Mermaid) tracing real request flows, including every error branch - A plain-English walkthrough explaining how the whole system fits together - A step-by-step local setup guide - A ranked list of concrete "good first tasks," each with exact file and line references, generated by a dedicated subagent that independently re-scans the codebase for well-scoped opportunities The target users are new developers (or new team members generally) joining an existing project, and engineering managers who currently spend hours manually onboarding each new hire. Instead of scheduling a walkthrough call or writing docs by hand, a lead can run this pipeline once and hand a new developer a complete, accurate onboarding kit, generated directly from the current state of the code, not from documentation that's months out of date. What makes this creative and unique is that it doesn't just summarize code, it builds a genuinely new artifact (a custom, project-agnostic Bob mode) that other teams can reuse on their own codebases, not just on this one. It also uses a subagent specifically to isolate the first-tasks analysis from the main documentation-generation context, which measurably improved output quality, the subagent-generated tasks were more concrete and better grounded in real code than the first pass. To test real effectiveness, I timed myself understanding a specific flow (booking a flight) two ways: reading raw source code cold, versus reading the generated walkthrough. For a single, well-written function, the two took roughly the same time, an honest result I kept rather than inflating. But the real value showed up in cross-cutting understanding: seeing how the REST and MCP layers share the same validation logic, how errors travel from backend checks to what a user sees on screen, and how the whole system fits together, the kind of knowledge that normally takes a new developer far longer to piece together by jumping between files and asking teammates. That's the actual problem this project solves.