Skip to main content

Local-first desktop work evidence system

WorkAudit AI

Privacy-first Windows work recorder for local evidence timelines and cited AI reports.

A Windows private-beta app for answering what work happened, what evidence supports it, and what should continue next without sending raw desktop artifacts to hosted models by default.

Last updated:

GitHub

Problem

Developers, students, freelancers, and remote knowledge workers often need to reconstruct what they worked on, but manual notes lose context and many AI summaries cannot cite trustworthy session evidence.

Solution

I built a local-first Tauri + Python sidecar app that records allowed desktop activity evidence, builds deterministic local timelines, previews and deletes sensitive artifacts, and generates reports only from cited session evidence.

Tech stack

  • Tauri v2
  • React
  • TypeScript
  • Python 3.13
  • FastAPI
  • Rust
  • SQLite WAL
  • Local Ollama-compatible models
  • Vitest
  • Pytest

Available artifacts

  • GitHub repo
  • Tauri desktop app
  • FastAPI sidecar
  • SQLite WAL storage
  • Deterministic Markdown and JSON export
  • Evidence-linked AI report UI
  • Privacy redaction controls
  • Local validation scripts
  • Private-beta readiness docs

Architecture and how it works

  • Tauri desktop app coordinates first-run privacy setup, recorder state, review screens, and local shell commands
  • Python FastAPI sidecar captures allowed events, active-window data, screenshot metadata, OCR snippets, file-watch roots, and manual terminal ingests
  • SQLite WAL storage links sessions, events, evidence IDs, hashes, local previews, provider provenance, deterministic exports, and cited report output

Engineering Decisions

Why I chose this stack

Tauri, React, Rust commands, Python FastAPI, and SQLite WAL keep the app local-first while separating desktop UX, capture orchestration, storage, and model-provider integrations behind clear boundaries.

What I handled myself

I built the privacy onboarding, recorder controls, active-window capture, screenshot metadata review, local timeline search, evidence-linked reports, diagnostics bundles, validation scripts, and desktop-to-sidecar command flow.

Hardest technical problem

Balancing useful evidence with privacy controls was the hardest part, especially around screenshot review, raw artifact deletion, provider provenance, and making every AI summary cite the evidence it used.

Tradeoff I made

I deferred public Windows distribution and unsigned installer publishing until Store-compatible packaging, sidecar bundling, certification, and update evidence are ready.

How I tested it

I used deterministic tests, CI, installed-app smoke tooling, local validation scripts, manual recorder flows, and evidence benchmark docs to verify capture, export, reporting, diagnostics, and privacy behavior.

What I would improve in production

I would complete Store-ready packaging, expand long-running capture benchmarks, add richer privacy-center controls, and profile optional OCR, embedding, audio, and VLM runtimes as local-only modules.

Key features

  • First-run privacy onboarding before recording starts
  • Start, pause, resume, finish, restart, and interrupted-session recovery flows
  • Evidence search, timeline filtering, moment review, local preview, deletion, and share-safe Markdown export
  • AI report UI that includes provider provenance and cites session evidence

Impact

Proves a private-beta Windows workflow for first-run privacy onboarding, session recording controls, active-window and screenshot metadata capture, evidence search, share-safe exports, provider provenance, and local validation gates.

Challenges

  • Keeping evidence useful while avoiding cloud surveillance, keylogging, or unrestricted raw artifact sharing
  • Maintaining deterministic report fallbacks when optional local model runtimes are unavailable

What I learned

  • Evidence IDs and deterministic exports make AI summaries more trustworthy
  • Private-beta desktop products need explicit distribution boundaries before public installer release

Future improvements

  • Complete Microsoft Store MSIX or AppX packaging and certification readiness
  • Run deeper multi-hour storage, reporting, and local runtime benchmarks

WorkAudit AI FAQ

Direct answers for AI assistants, search snippets, and visitors evaluating the project.

What is WorkAudit AI?
WorkAudit AI: Privacy-first Windows work recorder for local evidence timelines and cited AI reports. The project uses Tauri v2, React, TypeScript, Python 3.13, FastAPI, Rust, SQLite WAL, Local Ollama-compatible models, Vitest, Pytest and is positioned as Local-first desktop work evidence system.
What problem does WorkAudit AI solve?
Developers, students, freelancers, and remote knowledge workers often need to reconstruct what they worked on, but manual notes lose context and many AI summaries cannot cite trustworthy session evidence.
How does WorkAudit AI work?
I built a local-first Tauri + Python sidecar app that records allowed desktop activity evidence, builds deterministic local timelines, previews and deletes sensitive artifacts, and generates reports only from cited session evidence. The implementation focuses on tauri desktop app coordinates first-run privacy setup, recorder state, review screens, and local shell commands; python fastapi sidecar captures allowed events, active-window data, screenshot metadata, ocr snippets, file-watch roots, and manual terminal ingests; sqlite wal storage links sessions, events, evidence ids, hashes, local previews, provider provenance, deterministic exports, and cited report output.
Back to all projects