Own project, live app, public MIT source
Flipper
Captured from the real desktop interface with synthetic listings, prepared extraction results, illustrative price histories and run records. Discover uses AI-generated product photos based on manufacturer references. Detail crops keep the evidence readable; no marketplace seller data is used.
Compare local used-computer listings with matching nationwide asking prices, and inspect the evidence behind each reference.
What I built.
I built the web interface, marketplace adapter, hardware interpretation, deterministic comparisons, SQLite persistence and private job controls.
The engineering decision.
Use optional AI to extract source-supported specifications; keep price arithmetic and comparable-offer rules in deterministic code. Show uncertainty and separate public browsing from private collection and settings.
The result.
The public app is read-only, with specification filters, comparison evidence and observed price history. The private workspace adds saved searches, bounded refreshes and explicitly enabled daily scans.
After release
What broke, and what changed.
On 20 October 2026, Chợ Tốt answered the hosted daily scan with HTTP 429 and a 24-hour Retry-After, which stopped the desktop and GPU refreshes partway. Nothing kept that deadline, so the next category or a restart could send requests during the pause.
The fix, deployed the same day, keeps the pause in SQLite, holds every marketplace request until it ends, and queues the interrupted refreshes with what is left of their original AI budgets.
It worked in production. The desktop refresh finished on 21 October, when the pause ended. The GPU refresh met a second 24-hour pause and finished on 22 October, using $0.08 of its $0.25 budget.
Evidence and scope
Source tests and production recovery
- React / TypeScript
- Node.js / SQLite
- LiteLLM
References compare advertised asking prices, not completed-sale values, physical condition or resale profit. The rules for comparing prices are in the repository's README.
The screenshots use synthetic listings, prepared extraction results, illustrative prices and run records in an isolated local instance. The Discover view includes AI-generated product photos based on manufacturer references. They show the interface; the illustrated recovery does not establish production execution. No marketplace seller data is included.
Source tests cover interpretation, pricing, provider budgets, scheduler recovery and API access boundaries. They do not establish provider availability or future scheduled execution.