buildlog

Engineering insights that compound

Capture what works. Measure whether it actually helped. Drop what didn't.

buildlog extracts decision patterns from your AI-assisted work and uses a Thompson Sampling contextual bandit to surface rules that reduce mistakes, then tracks whether they did. The feedback loop is statistical, not vibes-based.

See it working

The Learning Loop E2E Trace walks through all 13 steps of the pipeline with explicit code citations. From installation through Thompson Sampling selection, gauntlet review, Bayesian bandit updates, emission pipeline, cross-domain discovery via qortex (a separate knowledge graph package in the ecosystem), and rule re-export back into buildlog.

Every claim on this page has a mechanical proof in that document.

Features

  • Capture structured journal entries from work sessions, mistakes included
  • Extract entries into reusable engineering rules with semantic deduplication
  • Select rules via Thompson Sampling bandit, automatically surfacing the most effective per context
  • Measure Repeated Mistake Rate (RMR) across experiments with statistical rigor
  • Review code through curated reviewer personas (Security Karen, Test Terrorist, Bragi)
  • Emit structured artifacts to ~/.buildlog/emissions/ for downstream systems
  • Bridge into qortex for cross-domain pattern discovery across projects
  • Integrate with Claude Code, Cursor, GitHub Copilot, Windsurf, and Continue.dev

Quick install

uv pip install buildlog        # or: pip install buildlog
buildlog init-mcp --global -y  # register globally for all projects

Quick start

buildlog init                  # scaffold a project
buildlog new my-feature        # capture a work session
buildlog distill               # extract patterns
buildlog skills                # deduplicate into rules
buildlog experiment start      # begin tracked session (bandit selects rules)
# ... work ...
buildlog experiment end        # close session
buildlog experiment report     # see the numbers

The pipeline

buildlog pipeline

Thompson Sampling closes the loop: rules are selected based on learned effectiveness, and feedback updates the model.

The ecosystem

buildlog is one piece. qortex is a separate package that provides the knowledge graph layer. buildlog emits structured artifacts; qortex ingests them, bridges rules to design pattern domains, discovers cross-domain relationships that no individual system can see, and projects new rules back. The Learning Loop E2E Trace documents exactly how this works at the code level (see sections 10-11).

buildlog emits    ──>  qortex ingests  ──>  graph discovers patterns
     ^                                              |
     └──────── qortex projects new rules  <─────────┘

Next steps