Alexa Toby
open sourceA custom Alexa skill that puts a real agent behind the Echo. Ask it anything and it answers through a hosted language model, with a web search tool for anything recent — stock wake word, stock speakers.
About this project
You cannot replace the assistant inside an Echo, because speech recognition and language understanding run in Amazon's cloud. A custom skill is the supported way to substitute the brain while keeping the wake word and Drop In intact, so this is a small self-hosted backend that does exactly that. The engineering is almost entirely one constraint: the reply must arrive within eight seconds, which buys exactly one tool round. If less than 2.5 seconds remain after the model decides to search, the search is skipped and it answers from memory instead — a deliberate trade that fails fast rather than being cut off mid-sentence. Search itself goes through DuckDuckGo's HTML endpoint with an instant-answer fallback and a five-minute cache, because it rate-limits by IP and returns a captcha page rather than an error. A second interaction model handles Hindi.