Backend Software
Engineering Case StudyInterview Platform
Three related tools for career evidence and interview practice, with explicit storage contracts and coding endpoints between them.
19 MCP tools · Markdown and SQLite backends
- My Contribution
- Built Recall-Seed, STARfolio, and Scroll; defined storage and endpoint contracts.
- Project Status
- Implemented · Active personal projects
STARfolioDesktop interviews + SQLite search
ScrollCoding endpoints + grading callbacks
The Problem
Interview practice needs concrete evidence about what someone has done. I wanted career notes, retrieval, and coding exercises to work across tools without coupling every feature to one desktop app.
Distinct Responsibilities
Recall-Seed is a C#/.NET MCP server. Its 19 tools retrieve and store career evidence, check citation identifiers, and seed coding challenges into Scroll. The calling model generates language; the server supplies source material and validation.
STARfolio is the Electron desktop surface for career notes and mock interviews. SQLite FTS5 and sqlite-vec support keyword and vector search.
Scroll is an implemented editor with sync server and agent-facing endpoints. Relative viewport anchoring keeps edits above the visible region from moving the reader’s screen.
Backend Decisions
One IExperienceStore interface supports Markdown-vault and SQLite implementations. Tools use that interface rather than knowing how each backend stores notes. The SQLite path uses WAL mode and a stable read/write contract.
Coding challenges cross a separate boundary: Recall-Seed provides the problem schema; Scroll hosts the editor and grader, then posts the verdict back with a correlation identifier.
Verification and Limits
Recall-Seed CI builds Release, runs xUnit tests, and exercises the stdio protocol from initialization through a tool call. Source-linked example runs show how thin evidence produces a gap question.
Citation checks verify that referenced identifiers exist. They are not a guarantee that every generated statement is true.
Scroll’s treap has differential property tests against a simpler reference structure. A targeted sweep of 20 hand-built mutants reached near-total detection; the separate whole-suite mutation run reported 86–95% on covered code. Those measurements assess test sensitivity, not production reliability.