CodeGuard learns your team's unwritten rules from merged PRs and applies them automatically. Powered by AST-level analysis, semantic search, and LLM reasoning.
From rapid ideation at 2am to validated architecture — Figma Make helped us think faster and iterate smarter.
Prototyped the two-pipeline concept (Evaluate + Commit) visually before writing any code. Aligned the team at 2am.
Architecture alignment at 2amCreated wireframes showing how an intern's commit flows through CodeGuard: from AST diff to guideline matching to review.
Pitched the idea to team visuallyKilled retroactive history mining early, saving 3+ hours. Realized attributes should only accrue from PRs going forward.
Threw away bad ideas earlyCreated the team breakdown visualization with interfaces and dependencies. Each person could see their scope at a glance.
Team aligned and shippingCodeGuard runs two distinct pipelines: one learns from merges, the other reviews every PR.
Click any module to explore its functions and internals.
positive + negative collections
pattern, reasoning, diff, confidence
touched layers, risky patterns, intent
LLM-first or rules-only modes
parse, diff_trees, extract bodies
cosine similarity, threshold, weight
on: pull_request, on: push (main)
Generalized attribute suppression
Follow the path — click each checkpoint to see how CodeGuard transforms a new contributor's journey.
The intern writes code that works, but puts business logic inside a route handler and uses except: to catch all errors. Tests pass, but team conventions are broken.
CodeGuard's Evaluate pipeline extracts AST diff via tree-sitter, embeds the changed functions, finds similar code, and flags issues.
The intern fixes the PR. When it merges, the Commit pipeline compares before/after, classifies the fix as positive, and stores the guideline.
Another team member makes a similar mistake. CodeGuard now has a stronger guideline from the first intern's experience. Review is faster and contextual.
The intern understands not just what to change, but why the system works this way. CodeGuard provided the unwritten rules that usually take months to absorb.
Four people, four roles, clean interfaces between each layer.
Memory infrastructure — ChromaDB, guideline schema, embedding wrapper, and repo intent loader.
Turns merged commits into reusable knowledge. Filters, classifies positive/negative, stores structured guidelines.
Runs on every PR. AST diffs, finds similar functions, retrieves guidelines, runs LLM compliance checks.
Wires everything into GitHub Actions, handles PR commenting, dismissal tracking, and the feedback UI.