ScoutDex

I play Pokémon GO, and the game keeps its real decisions behind spreadsheets — IVs, movesets, league rankings, which raids are actually worth it this week. The answers are out there, but finding them means tabbing between half a dozen community sites in the middle of a battle.

ScoutDex is the scout I wanted. Ask it anything — “is 15/14/15 good for Umbreon in Great League?”, “what should my Lucario run?” — and it answers in a sentence, pulled from current game data instead of a stale wiki page.

How it’s built — You reach for ScoutDex mid-battle, so latency drove most of the design — a slow answer is a useless one. The core decision was to ground every answer in real data instead of trusting the model’s memory: it parses your question into a structured lookup, pulls the exact numbers from a current copy of the game’s data, and hands those to the language model only to phrase. That retrieval step is what keeps it from inventing stats. On top of that it caches answers it has already generated, so repeat questions (“best moveset for X”) come back instantly instead of hitting the model again, and it streams the reply a word at a time so you’re reading before the whole answer is done.

The challenge that surprised me was the model provider. I’d left it on automatic routing — letting it pick a service per request — and that alone was adding almost a second to every answer; pinning it to a single provider fixed it. A scheduled job refreshes the game data on its own, so the numbers stay current without me touching anything. Live at scoutdex.io.