On-device extraction, encrypted storage, and a tightly scoped AI proxy.
The challenge
Most consumer-intelligence products centralize raw personal data on a server and call that insight. Payback needed to produce useful analysis without reproducing the same surveillance pattern it was supposed to critique.
How we built it
I moved extraction, normalization, and checkpointed long-running work onto the device with ZIP parsing, media agents, resumable background jobs, and encrypted SQLite storage. The backend is intentionally narrow: it verifies Google OAuth tokens, rate-limits aggressively, supports dual-key Gemini failover, and only handles derived category sets instead of raw exports.
What shipped
Payback turns Google and Meta exports into 135 behavioral categories across 10 pillars while keeping raw source data local. The interface, documentation, and backend all reinforce the same product promise: insight without surrendering the underlying archive.
Outcomes
- Kept raw Google Takeout and Meta export data on-device using AES-256-GCM encrypted SQLite.
- Normalized behavior into 135 categories across 10 pillars for consistent downstream analysis.
- Built resumable background processing plus a hardened AI proxy with OAuth verification, rate limits, timeout control, and API-key failover.