Voice understanding pipeline
Transcription is corrected, classified and transformed into structured intent before downstream tools or storage are invoked.
Voice-first agent system
VoiceVault
Voice becomes structured memory, actions and an agent conversation.
VoiceVault is a voice-first web application that turns spoken input into searchable notes, tasks, reminders, decisions and contextual agent actions. It combines transcription, structured extraction, conversational memory and tool execution rather than stopping at speech-to-text.
live intent
“Remind me to send the proposal after tomorrow's call.”
REMINDERTOMORROWAFTER CALLCONFIDENCE 0.96
Private product repository. Architecture facts are derived from 46 route handlers and the implemented AI, voice, memory and tool modules.
46
API routes
44
AI & voice modules
100+
voice intents
1
unified action layer
System anatomy
Complexity broken into reliable boundaries.
Conversational memory
Multi-turn context connects references, prior notes and the current action so follow-up commands do not restart from zero.
Tool-routed actions
A capability layer covers note operations, reminders, search, summaries, pattern analysis and suggestions with confirmation policies for sensitive actions.
Searchable personal knowledge
Structured extraction and semantic retrieval make voice history queryable as memory, not just a folder of audio files.
Engineering judgment
The decisions behind the build.
Seniority is not the number of technologies in a project. It is the ability to identify failure modes early and choose boundaries that keep the system understandable after launch.
Transcription is only stage one
The product value comes from turning speech into reliable structure and actions, not from producing a text blob.
Fast paths need safe undo
Low-risk voice commands can execute immediately while confirmation policies and an undo layer protect consequential actions.
Memory must remain useful
Context selection and semantic retrieval keep relevant history available without continually expanding the model context window.
Production stack
Chosen for the system.
Next.jsReactTypeScriptOpenAIWhisperSupabasePostgreSQLEmbeddingsNextAuthVitestPWA