Style Distiller

DR-001: Local-First Architecture

Date: 2026-10-15

Status: Active

Decision

All chat data analysis runs locally on the user's machine. No cloud backend, no data upload, no server-side processing.

Context

WeChat chat history contains highly personal data: conversations with friends, family, colleagues, romantic partners. Message content reveals:

  • Social relationships and network structure
  • Communication patterns and emotional tone
  • Code-switching behavior (language identity)
  • Timestamps (daily routines, sleep patterns)
  • Contact names and organizational affiliations

Sending this data to a cloud service would require:

  • User trust that we handle data responsibly
  • Infrastructure to secure data in transit and at rest
  • Compliance with data protection regulations (GDPR, CCPA, Australian Privacy Principles)
  • Ongoing operational costs and maintenance
  • Clear data retention and deletion policies

Options Considered

Option A: Cloud-based SaaS (rejected)

  • User uploads WeChat database to our server
  • Analysis runs server-side (faster, scalable)
  • Results stored in user account
  • Rejected: Privacy risk too high, operational complexity, user trust barrier

Option B: Local-first with optional cloud sync (rejected)

  • Analysis runs locally by default
  • Optional cloud backup of generated profiles (not raw chat data)
  • Rejected: Still requires backend infrastructure, sync complexity, partial privacy guarantee

Option C: Pure local-first (chosen)

  • Python CLI tool runs entirely on user's machine
  • No network calls except optional BYOK chatbot (user's own API key)
  • Chat data written to git-ignored folders, never committed
  • Demo site uses synthetic personas only, no data collection

Why

Privacy: User retains full control. Data never leaves their machine. No trust required.

Simplicity: No backend to build, secure, or pay for. No auth, no database, no ops.

Transparency: Open-source Python pipeline. Users can audit every line.

Cost: Free to run indefinitely. No usage limits, no subscription.

Compliance: No data processing = no regulatory burden.

What Happened

The local-first architecture worked. The Python tool successfully analyzes WeChat databases on macOS with zero network dependencies (except optional chatbot). Users who tried it appreciated the privacy guarantee.

Trade-offs accepted:

  • Installation friction (requires Python 3.12+, chatlog tool, WeChat desktop)
  • No cross-device sync
  • No hosted service for non-technical users
  • Limited to desktop platforms (macOS, Linux)

Metrics (as of 2026-10-15, based on test runs with synthetic data):

  • Analysis speed: ~150 messages/second on M2 MacBook Air
  • Memory footprint: <100 MB for 10k messages
  • Dependencies: stdlib only (optional: jieba for Chinese, anthropic for chatbot)

What I'd Change

If building this as a commercial product for non-technical users, I'd consider:

  • Browser-based tool: WASM-compiled Python or pure JS implementation that runs in the browser. Still local-first, but zero installation.
  • Mobile support: iOS/Android apps with local processing. Harder to extract WeChat data on mobile (app sandboxing), but possible with export features.
  • Differential privacy: If any aggregate statistics were shared (e.g., "average message length across 1000 users"), add noise via differential privacy to prevent re-identification.

For this personal project, pure local CLI + demo site was the right trade-off.