Offline-first Android app for inventory and shopping management — built for one specific person who uses it every day at work and at the supermarket. No backend, no account, no internet required. Data belongs to the user by architecture, not by policy. The name comes from monopsônio popular — the consumer, collectively, as the dominant buyer. This is the first layer of that idea made functional.
- Full product ownership — schema design (V1→V10 versioned migrations), UI architecture, UX driven by observed behavior. Real users, real feedback loop, real consequences when things break.
- Performance engineering, three separate wins — New Architecture/JSI eliminated the JSON bridge (cold start 3s → 1s; ~90% of total gains from this alone); WAL mode + explicit PRAGMAs eliminated lock overhead (repeat-visit load 900ms → 33ms); profiling identified a single query as bottleneck (1287ms → 5ms after indexes in V8 migration).
- Source of truth follows observation — during v1.7 development, caught a fundamental data model flaw: the app was storing a derived value (price per unit) as canonical. User typed R$14,00 and got R$14,0000000003 back. The rounding error was a symptom; the real fix was epistemics. Corrected before shipping.
- Spec-driven AI-assisted development — complex features fully specced before implementation. AI proposes, human decides, human reviews every diff, human owns the architecture. Faster iteration without surrendering design ownership.
End-to-end market price intelligence for Rio de Janeiro supermarkets — scraper → normalization → dual-layer storage → API → full product layer. Tracks 18,000+ prices daily. Built in two weeks during a financial crunch, then kept growing: generics, cross-store comparison, shopping lists, fuzzy paste import, Monopop export. The loop is closed — know your stock, know the market, build your list, buy smarter.
- Per-store concurrency control — asyncio.Semaphore scoped per store independently, not globally. A global semaphore lets one slow store (VTEX averages 5–6s per call, spikes to 20–30s) consume all slots and starve the others. Per-store isolation contains the damage. Result: full scrape cycle 3 hours → 6 minutes (30×).
- Dual-layer storage — reactive SQLite cache (4h TTL, low-latency search) + PostgreSQL time-series (90-day price history). Two different problems — search latency and trend data — each solved with the right tool. One database wouldn't handle both correctly.
- Full product layer built on the data pipeline — normalized generic pages group cross-store variants by brand and package size with price-per-unit comparison. Shopping lists support paste-driven fuzzy import (bigram + Levenshtein scoring with a mid-confidence review step before committing), variant price pinning, and one-click export as Monopop-compatible JSON. Search is stateless SSR; lists are client-side — two different statefulness requirements handled separately rather than forcing a single model.
- Self-healing retry pipeline — GitHub Actions cron logs every scrape result to a table. The hourly retry job queries for failures in the last 24h with no later success and re-enqueues only those (term, store) pairs. No dead-letter queue, no fixed retry counter. Idempotent writes throughout.
Internal platform used by prosecutors across Rio de Janeiro state. SPA awarded 2nd place — 2021 CNMP National Innovation Award. Led architecture, CI/CD, and a design system that scaled across teams.
Client work running in parallel with institutional employment since 2016 — landing pages, internal tools, web applications. Full ownership at the sharp end: reading what the client actually needs, writing the spec they couldn't, delivering it, and making sure the thing still works six months later.