BACK TO INDEX
CASE-2026-018 // LEARNGRAPH
UNCLASSIFIEDDeveloper Tools / Learning Analytics

LearnGraph

STATUS:ACTIVE
RISK:LOW

EXECUTIVE SUMMARY

LearnGraph is a lightweight Windows learning tracker built around the same capture idea as Laya Inbox but optimised for a different job: sustained, visible progress. A global hotkey (Ctrl+Shift+Space) logs what the user learned, a locally hosted model classifies each entry into a user-defined taxonomy, and the result is rendered as GitHub-style contribution heatmaps — one per domain — with a review queue for unclassified or low-confidence entries. The heatmap is the product; the dashboard is secondary.

PROBLEM STATEMENT

Learning progress is usually tracked in tools built for tracking tasks, which measure completion rather than retention and produce a to-do list rather than a picture of sustained effort. Habit trackers record that a session happened but not what was covered. The result is that learners cannot answer the only question that matters — am I actually making progress in the areas I care about — because the evidence exists but has never been visualised in a form the eye can read at a glance.

INVESTIGATIVE FINDINGS

  • ▸Contribution heatmaps communicate sustained effort faster than any aggregate statistic
  • ▸Per-domain heatmaps are necessary because time spent is not progress unless the domain is separated
  • ▸Classification confidence must be surfaced, not hidden, or users will not trust the map
  • ▸A local model is fast enough that capture stays interactive, which is what keeps the habit alive
  • ▸Electron was rejected here: a native binary that launches instantly is a prerequisite for a global-hotkey tool

BEHAVIORAL ANALYSIS

The product bets that visibility is the retention mechanism: a heatmap that fills in is itself the reward, so no notification system is required. Classification confidence is shown rather than hidden, so a wrong entry is visibly wrong and easy to correct, which keeps the map honest enough to stay motivating over months.

TECHNICAL EXECUTION

The tracker is a Tauri 2 application with a Rust core and a React 19 + TypeScript frontend in strict mode. The Rust side owns the hotkey registration (Ctrl+Shift+Space), the SQLite layer via bundled rusqlite, and the HTTP client (ureq) used to reach the local model service. Persistence is deliberately persist-first: an entry is written to SQLite before any remote classification call is attempted, with a retry path in the pipeline module so a failed classification never loses a captured entry. Classification runs through the daemon in a two-pass process that was tuned against live traffic, and `src-tauri/src/laya.rs` holds the classification logic while `pipeline.rs` owns the retry behaviour. Eight Rust tests pass, including a live-daemon smoke test. The release executable is 7.6 MB and idles at roughly 33 MB of memory. No Electron, no Node server, and no charting library — heatmaps are drawn natively.

SYSTEM ARCHITECTURE

Global Hotkey (Ctrl+Shift+Space) → Rust Hotkey Handler (Tauri 2) → Persist-First SQLite Write (rusqlite, bundled) → Two-Pass Laya Classification (laya.rs) with Retry Queue (pipeline.rs) → Taxonomy Assignment → Per-Domain Heatmap Rendering (native, no chart library) → Review Queue for Low-Confidence or Unclassified Entries → Settings and Taxonomy Manager.

OPERATIONAL FEATURES

  • ✓Global hotkey capture from any application on Windows
  • ✓Persist-first write path so no entry is ever lost to a failed model call
  • ✓Two-pass local classification against a user-defined taxonomy
  • ✓Retry queue for classification failures
  • ✓GitHub-style contribution heatmaps rendered per learning domain
  • ✓Review queue surfacing low-confidence and unclassified entries
  • ✓Built-in taxonomy manager for defining and editing domains
  • ✓Settings panel
  • ✓Release binary of 7.6 MB idling at roughly 33 MB
  • ✓Eight Rust tests passing, including a live-daemon smoke test

TECHNOLOGICAL ASSETS

Tauri 2RustReact 19TypeScriptrusqlite (bundled SQLite)ureqLocal GPU Inference (Laya)

CURRENT STATUS

V1 core complete and verified. Hotkey capture, persist-first persistence, two-pass classification tuned against live traffic, multi-domain heatmaps, review queue, taxonomy manager, and settings are all implemented, with eight Rust tests passing including the live-daemon smoke test. Current work focuses on export formats and richer review workflows.

CONCLUSION

LearnGraph demonstrates that the smallest useful version of a tracking tool is a capture hotkey, a durable write, and one visualisation that rewards the user. Everything else — dashboards, statistics, notifications — is optional once the heatmap is trustworthy, and the persist-first ordering is what makes the map trustworthy enough to keep looking at.

EVIDENCE LINKS

REPOSITORY

CASE METADATA

OPENED:2026-10-01
ID:CF-2026-018

TAGS

Local-FirstDeveloper ToolsTauriRustLearning AnalyticsVisualizationHeatmapHabit TrackingWindowsSQLite
SESSION ACTIVELAST SYNC: --:-- --/TRACE:HOME/CASES/CF-2026-018//ACTIVE[ESC] to exit contextreturning may invalidate findings
○ STANDARDCLEARANCE: LEVEL 2