kReative Labs
Back to kReative Labs

Building 2026

Aura

A personal AI clone. Clone your voice from a handful of audio samples, feed it a private knowledge base, then ask it questions and hear the answers back in your own voice.

View source

The idea

A voice, a memory, and a way to ask it things

Three separate problems have to work together for a persona to feel like more than a chatbot with a filter on it: the voice has to sound like you, the answers have to be grounded in what you actually know rather than what a base model guesses, and both have to run without demanding infrastructure a single person would not normally stand up.

Aura keeps each piece as small as the problem allows. Voice cloning is one API call to a provider built for it. Retrieval is BM25 computed in-process against SQLite, no vector database or embedding key required, because a personal knowledge base does not need one. A ghost mode setting instructs the persona to withhold identifying details when that matters more than completeness.

Architecture

Three inputs, one persona

Aura architecture Audio samples go to the ElevenLabs Instant Voice Clone endpoint and produce a voice ID, auto-assigned to the persona. Documents are chunked on paragraph boundaries and stored in SQLite, with BM25 retrieval computed in-process at query time. A question triggers retrieval, then a system prompt built from persona configuration, retrieved chunks, and standing rules is sent to Claude, which streams a cited reply over server-sent events. The reply can optionally be rendered through the cloned voice for spoken output. Audio samples recorded or uploaded ElevenLabs instant voice clone -> voice ID Documents PDF, text, code SQLite + BM25 chunked ~1200 chars, retrieval in-process Your question typed in Persona config + retrieved chunks + standing rules -> Claude, streamed reply cited inline, ghost mode aware Spoken reply in cloned voice Text streamed, cited [1] [2] optional
Voice and knowledge are independent inputs that only meet inside the persona: a question is answered from retrieved text first, and only spoken through the cloned voice if that step is asked for.

The app runs with no API keys configured. Anything left as a placeholder disables that feature cleanly rather than crashing; "Connect Social Profiles" and calendar sync are documented stubs with no ingest pipeline behind them yet.

Detail

At a glance

VoiceElevenLabs Instant Voice Clone. First cloned voice auto-assigned to the persona; generated audio served back from local storage.
KnowledgePDFs and text chunked on paragraph boundaries (~1200 chars), stored in SQLite. BM25 retrieval computed in-process, no vector store or embedding key.
PersonaSystem prompt built from domain, tone, verbosity, expert mode, and ghost mode, plus retrieved chunks and standing rules. Claude Opus 5 streams the reply over SSE with inline citations.
Ghost modeInstructs the persona to withhold identifying details (employers, clients, locations); generated audio is tagged accordingly.
StackNode/Express, SQLite via node:sqlite, vanilla JS frontend.
StageBuilding. Runs locally; not yet deployed.