HoguEscape

Inspiration

Marketplace listings make the same product look different through inconsistent titles, model names, discounts, and metadata. HoguEscape helps Korean shoppers answer one question before buying: is this price actually good, or should I wait?

What it does

HoguEscape combines live offers from multiple marketplaces, groups exact models and configurations, and explains a transparent purchase-risk score. Users can compare sellers, inspect price history, browse verified deals, and create price alerts on web and mobile.

What we built during OpenAI Build Week

HoguEscape existed before the event. This submission focuses on meaningful extensions built with Codex and GPT-5.6 after July 13, 2026: recovering Coupang acquisition after its API limit contract changed, shared quota accounting and fail-closed diagnostics, registry-driven model grouping for tablets and robot vacuums, stable daily price history, repaired analytics attribution, and end-to-end affiliate-click verification. Dated commits and Codex sessions separate this work from the pre-existing product.

How we built it

Codex worked across the FastAPI backend, Next.js web app, and React Native client. GPT-5.6 helped convert production symptoms into testable hypotheses, regression tests, and focused fixes. We reproduced each symptom, inspected real provider payloads, added a regression, implemented the smallest shared fix, and verified tests, CI, deployment health, and the live flow separately.

Challenges

The hardest bug returned HTTP 200 with an application-level error when the requested result limit exceeded the provider's new maximum. AI initially changed caches, filters, and retries because those were visible in code. Comparing live responses at limits 10, 20, 50, and 100 revealed the real contract change.

Marketplace identity was equally difficult. Aggressive fuzzy matching can merge different capacities, generations, or variants, so our registry-driven grouping prefers unknown over a false match.

Accomplishments

We now have production-safe quota handling, exact-model comparison, transparent score deferral when evidence is insufficient, stable price history, alerts, affiliate attribution, and regression coverage across backend and web flows.

What we learned

AI coding agents are strongest when paired with operational evidence. Passing tests and HTTP 200 responses are not enough; provider contracts, cache lineage, deployment state, and visible behavior must be verified independently.

What's next

We will expand exact-model coverage, improve cold-search latency, monitor price-history integrity, and make score explanations easier for non-technical shoppers.

Built With

• gpt-5.6