Outside

Outside was built by developer defiantlydefying for the Hacktoberfest open-source AI challenge (Touch Grass theme).

Outside's core design is to remember user refusals: when a user says a certain practice is not suitable for them, it permanently stops that practice for the entire session rather than just skipping one round.

Outside deliberately foregoes mechanisms designed to bring users back, such as streak tracking, push notifications, feeds, achievement badges, and accounts.

Outside's model is explicitly instructed to never say 'close your eyes,' and a regex validator rejects any out-of-bounds responses.

Outside does not request location permissions and has no location entries in its Info.plist.

Outside's conversation memory expires after ten minutes on the backend, and clicking Finish immediately deletes the conversation and all audio segments.

Outside's technical stack is: Expo/React Native for audio recording, an Express backend on Render, ElevenLabs Scribe for transcription, a Mastra workflow invoking the Gemma model, and ElevenLabs TTS for speech synthesis.

Gemma returns a structured object containing text, mode, quietSeconds, and practice fields, with the model driving UI rendering rather than merely populating a text box.

validateReply limits each spoken response to 65 words per turn and directly invalidates any response matching specific dangerous-phrasing regexes (such as hold your breath, close your eyes, cure, diagnos, etc.).

Outside's system prompt treats user speech as untrusted descriptive data rather than instructions, in order to resist prompt injection attacks.

Outside uses an adaptive noise-floor VAD strategy: it calibrates the noise floor from the first 600 milliseconds of each listening window, counts only audio above max(-40dB, floor+10dB) as speech, and ends a turn after 1.6 seconds of silence following at least 0.5 seconds of speech.

The developer's rationale for choosing open-weight Gemma over closed-source hosted models is behavioral auditability, making the entire chain—including the system prompt, validators, and word count limits—auditable.

The developer's actual production deployment uses hosted Gemma inference and proprietary ElevenLabs speech, rather than a fully offline or open-source technology stack.

Outside is an open-source meditation app that provides voice-guided outdoor meditation sessions conducted with the eyes fully open.

The project was built almost entirely in collaboration with the coding agent Codex, and AGENTS.md specifies not to trust training data knowledge about Expo and to instead fetch the corresponding SDK versioned documentation first.