DealScout

Problem

Watching for restocks and price errors on high-demand gear (GPUs, consoles, LEGO) across multiple stores is a manual, easy-to-miss chore. DealScout is a compliance-first deal-intelligence platform: it keeps price and availability history, detects restocks and unusually large drops, and scores the result — read-only by design, so a human still approves every purchase.

Approach

The stack is Python 3.12 with FastAPI and HTMX for a read-only dashboard, SQLAlchemy and Alembic over SQLite in WAL mode, and APScheduler for polling. Official APIs and affiliate feeds come first (Best Buy, eBay, Keepa), with scraping only where a source's terms allow it; every outbound request passes a shared rate limiter and cache, and every fetch, detection, and alert is audit-logged. No cart, checkout, or payment code path exists, and CAPTCHAs, queues, and bot protections are never bypassed.

Currently building: Week-1 and Week-2 have both shipped — ingest, scoring, delivery (push, email digest, budget enforcement), a read-only dashboard, and the Keepa and eBay Browse adapters that were still in the backlog last check. The MVP itself is now complete. The newest thread is a pivot the commit log calls the LLM-wrapper: turning DealScout into a conversational assistant with a provider-agnostic tool-calling layer (Claude/OpenAI/local), a streaming chat UI, session login, and injection-hardened agent instructions — landed this week, with a graceful-degradation fix already following it.

Outcome

The repo is private while it's early. This page will link out once there's something safe to share publicly.