Methods & Design Decisions

Original Tool: Analyzes real WeChat chat history extracted via the chatlog tool (MIT licensed). Requires WeChat desktop app on macOS/Windows and user consent to access their own SQLite chat database.

Demo Site: Uses 6 synthetic personas generated by scripts/generate_synthetic_personas.py. Each persona has 150+ synthetic messages with controlled linguistic features. No real user data is used, committed, or deployed.

Sample Sizes: Original tool works with any chat history size (tested 100–10,000 messages). Synthetic personas each have n=80–150 turns (a "turn" is one continuous sending block, which may contain multiple bubbles).

Pipeline: Extract → Clean → Analyze → Distill

1. Extract: Read messages from WeChat SQLite database (or synthetic message list). Fields: sender, timestamp, content, contact ID, is_group.
2. Clean: Group consecutive messages from the same sender into turns. Filter out system messages, media-only bubbles (images, stickers without text).
3. Analyze: For each turn, compute 12+ linguistic features:
   • Message shape: Length distribution (mean, median, p25–p90), bubbles per turn, punctuation habits
   • Expressiveness: Emoji frequency (Unicode + WeChat faces), laughter tokens (哈哈, lol, 233, hhh)
   • Language: Code-switching ratio (CJK chars vs Latin words), catchphrases (top n-grams), openers/closers
   • Temporal: Active hours (UTC), per-contact tone (avg length by recipient)
4. Distill: Map features to LLM instructions. E.g., "median 18 chars, p90 36 chars" → "Keep replies short: median ~18 characters, rarely over 36". Generate persona_prompt.txt (system prompt) and profile.md (human-readable summary).

Evaluation (Career-Aligned Upgrade)

• Inter-persona variance: Compute mean, variance, and 95% bootstrap CIs for each feature across personas. Verifies that personas differ meaningfully.
• Pairwise similarity: Cosine similarity in 6D feature space. Lower values = more distinct persona pairs.
• Distinctiveness: For each persona, compute 1 - average similarity to all others. Higher = more unique profile.
• Feature correlation: Pearson correlation matrix. Identifies which features co-vary (e.g., emoji rate vs code-switching).

All confidence intervals computed with bootstrap percentile method (1000 iterations, fixed seeds 42–47 for reproducibility). Wilson score intervals used for proportions.

What AI Does: Optional BYOK (Bring Your Own Key) features allow visitors to test persona prompts and tune them interactively. AI is used for:

• Persona testing: Generate sample responses using a persona prompt to verify linguistic consistency.
• Prompt tuning: Adjust persona parameters (e.g., "make more formal") and compare before/after outputs.
• Persona matching: Match a visitor's self-reported texting habits to the closest synthetic persona.

What AI Never Does:

• Access your real chat data (the demo uses synthetic personas only)
• Store or transmit your API key to our servers (keys stay in your browser)
• Make claims about your personality, mental state, or identity
• Generate content for deployment without human review

Data Sent to Providers: When you use BYOK features, your browser sends:

• To Anthropic/OpenAI: Your prompt + persona prompt + test input (e.g., "How was your day?"). Your API key authenticates the request. Data is subject to the provider's privacy policy.
• NOT sent: Your API key is never sent to our server. All calls go directly from your browser to the AI provider (CORS enabled via anthropic-dangerous-direct-browser-access: true header for Anthropic, standard OpenAI API for OpenAI).

Human-in-the-Loop: All AI outputs are labeled "AI-generated". Every interaction is logged in an audit trail (stored locally in IndexedDB) with your decision: accepted / edited / rejected.

Audit Trail: View all AI interactions at /ai-log. Export as JSON or CSV. Clear anytime.