"Your spending tells a story. Your calendar, conversations, and emotions tell the rest."

Built for the Data Portability Hackathon 2026, Track 3: Personal Data, Personal Value.

Your bank statement shows what you spent. LifeLedger shows why.

By fusing exported data across transactions, calendar events, AI conversation history, emails, and lifelog streams, LifeLedger surfaces behavioral patterns that are invisible in any single source. The core insight: financial behavior is driven by stress, emotional state, and calendar pressure — not just income and expenses.

This isn't a dashboard that reformats your bank CSV. It's a cross-source intelligence engine that detects non-obvious patterns like:

• "Your discretionary spending spikes 40% in the week after high-stress calendar periods" — by correlating Google Calendar meeting density with transaction amounts across multiple alignment windows
• "You worry most about money in weeks where you also spend the most" — by scanning ChatGPT/Claude conversation exports for anxiety keywords and overlaying them with weekly spend
• "You're undercharging by $15/hr based on your invoice emails vs. calendar hours" — by extracting dollar amounts from email text and inferring hours from trailing calendar project blocks
• "62% of your spending happens within 3 days of payday" — by detecting income deposits and measuring the post-payday spending concentration

None of these insights exist in any single data export. They only emerge when you connect the dots across sources — which is exactly what data portability makes possible.

Most personal finance tools read one data source and display it. LifeLedger merges 5+ data sources into a unified behavioral timeline and runs correlation analysis across them. Calendar stress scores are compared against spending patterns. AI conversation sentiment is overlaid with financial behavior. Email invoice amounts are cross-referenced with calendar hours.

Every insight LifeLedger surfaces requires connecting at least two data sources:

Users can upload their own bank CSV (Chase, BofA, Amex, Mint), Google Calendar ICS, and ChatGPT/Claude conversation exports. Everything is processed locally. The AI chat supports Bring Your Own Key (Groq, OpenRouter, or OpenAI) — keys are session-only and never stored.

The narrative chat doesn't hallucinate. It receives only the precomputed insight JSON (never raw data) and answers questions strictly grounded in what the analysis actually found.

Two synthetic personas demonstrate the engine's capabilities:

Jordan Lee (p01) — Burnout + home savings goal. Shows stress-spend correlation, spike week evidence with specific transactions and calendar events, worry timeline, subscription creep, and savings velocity tracking.

Theo Nakamura (p05) — ADHD + freelance. Shows invoice rate risk detection (undercharging at implied $50/hr vs. $65/hr Austin baseline), anxiety themes tied to client stress, and spending pattern analysis.

lifeledger/
├── web/                                    # React + Express web application
│   ├── client/src/
│   │   ├── pages/
│   │   │   ├── Welcome.tsx                 # Landing page with demo/upload routes
│   │   │   ├── Dashboard.tsx               # Demo persona analysis view
│   │   │   └── YourData.tsx                # User upload + analysis view
│   │   ├── components/dashboard/
│   │   │   ├── KPICards.tsx                 # Dynamic KPI grid (only renders real data)
│   │   │   ├── TimelineChart.tsx            # Stress x Spend correlation timeline
│   │   │   ├── WorryTimeline.tsx            # AI conversation worry x spending overlay
│   │   │   ├── SpikeEvidence.tsx            # Spike week drill-down with transactions + events
│   │   │   ├── SubscriptionPanel.tsx        # Recurring charge detection
│   │   │   ├── PostPaydaySurge.tsx          # Post-income spending concentration
│   │   │   ├── DayOfWeekChart.tsx           # Day-of-week spending patterns
│   │   │   ├── BehavioralInsights.tsx       # Theme extraction + rate risk + savings
│   │   │   ├── StrengthsWeaknesses.tsx      # Data-driven strengths/weaknesses summary
│   │   │   ├── DataUploadSection.tsx        # Multi-file drag-drop upload
│   │   │   ├── UserContextForm.tsx          # Income/savings/debt context form
│   │   │   ├── ApiKeyConfig.tsx             # BYOK API key input (session-only)
│   │   │   ├── GroundedChat.tsx             # AI chat for demo personas
│   │   │   └── GroundedChatUpload.tsx       # AI chat for user-uploaded data
│   │   └── lib/api.ts                      # API client + types
│   └── server/index.ts                     # Express API (5 endpoints)
├── src/
│   ├── loaders/
│   │   ├── persona_loader.py               # JSON/JSONL normalization → unified timeline
│   │   └── upload_parser.py                # Bank CSV, ICS, ChatGPT/Claude JSON/ZIP parsing
│   ├── features/
│   │   ├── stress_scorer.py                # Calendar → daily stress scores (meeting density + deadlines)
│   │   ├── spend_tagger.py                 # 9-category discretionary spend tagging
│   │   ├── correlation.py                  # Multi-alignment correlation + 3-tier spike detection
│   │   └── resilience_model.py             # Financial resilience metrics (stability, volatility, runway)
│   ├── insights/
│   │   ├── insight_engine.py               # End-to-end insight computation + schema validation
│   │   └── narrative_gen.py                # LLM narrative generation (Groq → OpenRouter → OpenAI)
├── scripts/
│   ├── process_upload.py                   # Bridge: stdin JSON → upload_parser → insights → stdout
│   ├── demo_dry_run.sh                     # 2-minute timed demo runbook
│   └── generate_demo_backups.py            # Backup panel data per persona
├── data/raw/persona_p{01,05}/              # Synthetic persona exports (10 files each)
├── outputs/
│   ├── insights_p01.json                   # Frozen demo cache — Jordan Lee
│   └── insights_p05.json                   # Frozen demo cache — Theo Nakamura
├── context/
│   └── docs/how_to_export_your_own_data.md # Guide for exporting your own data
└── schemas/DATASET_SCHEMA.md               # Full schema reference

• Python 3.12+ with pip
• Node.js 20+

git clone https://github.com/RogueTex/LifeLedger.git
cd LifeLedger
python -m pip install -r requirements.txt
cd web && npm install && cd ..

cp .env.example .env
# Add one of:
#   GROQ_API_KEY=gsk_...        (recommended — free, fast)
#   OPENROUTER_API_KEY=sk-or-...
#   OPENAI_API_KEY=sk-...

Or skip this — users can enter their own API key directly in the chat UI (BYOK).

# Terminal 1: API server
cd web && npx tsx server/index.ts

# Terminal 2: Frontend
cd web && npx vite --host

Open http://localhost:5173. The API runs on :5000, and Vite proxies /api requests automatically.

• View Demo — Pre-analyzed synthetic personas (Jordan Lee, Theo Nakamura)
• Analyze My Data — Upload your own bank CSV, Google Calendar ICS, or ChatGPT/Claude exports
• Ask AI — Open the chat sidebar and ask questions about the insights (requires an API key via .env or BYOK)

Calendar events are scored for stress (meeting count, deadline keywords, packed schedules), smoothed over 7 days, and aggregated weekly. The engine tests 4 alignment strategies (same-week smooth, prior-week smooth, same-week raw, prior-week raw) and selects the strongest valid signal. Spike weeks are detected with a 3-tier threshold system that combines spending anomalies with stress levels, and each spike includes the top transactions and calendar events that explain it.

AI conversation exports (ChatGPT, Claude) are scanned for 18 worry-related keywords across financial and emotional categories. Mentions are grouped by week and overlaid with discretionary spending to reveal when anxiety and spending move together — a pattern invisible in either source alone.

Email text is scanned for invoice/payment signals and dollar amounts. When explicit hours aren't stated, the engine infers them from trailing 28-day calendar project blocks. If the implied hourly rate falls below the $65/hr market baseline, it flags undercharging risk and estimates the monthly leakage.

Users upload files through a drag-and-drop interface. The system auto-classifies by extension, merges multiple files of the same type, and runs the full insight pipeline. An optional context form (income, savings goals, debt) enriches the analysis with goal-aware projections.

Every insight includes: id, title, finding, evidence[], dollar_impact, what_this_means, recommended_next_actions[].

Up to 11 insight types are computed (all data-contingent — only generated when the data supports them):

• stress_spend_correlation — with weekly_series[], spike_weeks[], correlation stats
• worry_timeline — with timeline[], total_worry_mentions, peak week
• subscription_creep — with subscriptions[], monthly_total
• expensive_day_of_week — with by_day{}, pct_above_average
• post_payday_surge — with surge_pct, post_payday_total
• months_to_goal — with savings velocity and estimation mode
• top_anxiety_themes — with top_themes[] from conversation + lexicon
• stress_category_shift — which discretionary categories spike on high-stress weeks (calendar + transactions)
• spending_velocity — first-half vs second-half spend pacing within each pay period
• recovery_spending — spending increase the week after high-stress periods (decompression pattern)
• invoice_rate_risk — with matches[], implied rates, leakage estimate (persona-specific)

LifeLedger exists because of data portability. Every insight requires data the user already has the right to export:

See context/docs/how_to_export_your_own_data.md for step-by-step export instructions.

• All demo data is 100% synthetic (pii_level: "synthetic")
• User-uploaded files are processed locally in the browser session — nothing is persisted
• The AI chat receives only precomputed insight JSON, never raw transaction data or files
• BYOK API keys are session-only and never written to disk
• Synthetic data will be deleted after March 31, 2026 per hackathon rules

Burnout + home savings. Primary demo: stress-spend correlation with spike evidence, worry timeline showing anxiety-spending overlap, subscription creep detection, savings velocity to $50K goal.

ADHD + freelance. Secondary demo: invoice rate risk detection (undercharging $15/hr below market), ADHD/client-stress anxiety themes, spending pattern analysis without traditional income structure.