How Kluely Works

Audio capture and transcription

Kluely captures microphone audio (or, in Live mode, browser-tab audio). An AudioWorklet resamples it to 16 kHz, mono, 16-bit PCM and splits it into 50-millisecond chunks. These stream over a WebSocket directly to the chosen real-time speech-to-text provider, selectable on the home screen. Whichever is active is wired behind one interface, so the rest of the pipeline is identical either way. Authentication uses a short-lived token minted server-side; the API key never reaches the browser. The provider returns transcript text as you speak and, through automatic turn detection, signals “end of turn” when a thought is complete.

Handoff to the language model

On end of turn, the client sends the finalized transcript to Kluely's own server, which passes it to Google Gemini. The transcription provider and Gemini do not communicate directly; the app routes the transcript between them.

Translation and validation pipeline

1. 1

   Generate answer (English). Gemini writes a concise interview answer (shown as "Suggested").

2. 2

   Simplify. The answer is rewritten into short, concrete statements Klingon can express.

3. 3

   Handle missing words. For concepts Klingon lacks, substitute the closest real word (salmon → fish); only if none exists, transliterate phonetically as a loanword.

4. 4

   Ground vocabulary. Look up verified Klingon words for each concept in the boQwI' lexicon; the model may use only those.

5. 5

   Translate. Gemini composes the Klingon under a grammar reference (word order, prefixes, suffixes).

6. 6

   Render script. Convert the Klingon to pIqaD via a fixed lookup table (not the model).

7. 7

   Validate morphology. Parse every word with yajwiz, a Klingon morphological analyzer, to confirm validity and extract morphemes.

8. 8

   Back-translate. Build the literal English gloss ("Literal") from the yajwiz parse, not from the model re-reading its own output.

9. 9

   Check meaning and retry. Compare the back-translation to the intended answer and verify possessives, subject/object agreement, and plural endings. Regenerate with a correction if wrong.

10. 10

    Flag confidence. Mark the result verified, or unverified if it cannot be confirmed after retries.

Warp drive

Warp drive reduces latency by skipping the expensive verification steps — morphological validation, meaning checks, and retries — while keeping vocabulary grounding and simplification. Output is faster but not independently validated, and is marked accordingly.