VectorVoice
A conversation workspace that keeps the decisions, actions, and source moments people need without building a permanent archive of every word.
Why I built it
Most meeting products are designed to capture and retain as much as possible. That creates a growing archive of sensitive audio and transcripts even when the durable value is a much smaller set of decisions, actions, and supporting moments.
I built VectorVoice around a different product model: preserve the useful outcome and the evidence required to trust it, while minimizing the speech data that survives processing.
The product problem
A generated summary is difficult to trust when a reader cannot open the exact moment behind a claim. Retaining complete recordings and transcripts indefinitely solves retrieval at the cost of unnecessary privacy exposure.
The challenge was to create a concise, navigable record whose important claims remain verifiable and whose external actions remain under human control.
How I approached it
Capture without a meeting bot across web and native surfaces, normalize provider output, and keep transcription separate from the intelligence workflow.
Delete temporary audio after transcription, keep the complete transcript only long enough to generate structured output, then retain the minimum cited moments supporting summaries, decisions, and actions.
Keep external actions reviewable: the system can propose a follow-up or workflow update, but a person approves and owns the commitment.
How the product works
- 01
Capture
Web, macOS, and iPhone clients send temporary audio through one authenticated product API; the watch remains a scoped companion surface.
- 02
Transcribe
A provider-neutral worker produces a versioned normalized transcript and verifies deletion of the temporary audio object.
- 03
Understand
A separate intelligence service validates structured summaries, decisions, actions, and their supporting source moments.
- 04
Minimize
The structured output and cited evidence are committed while the complete transcript is deleted in the same database transaction.
Decisions that make it different
Evidence is a product feature
Every generated claim should open to the source moment that supports it, turning trust into a navigable interface rather than a disclaimer.
Provider-neutral business data
Speech and LLM adapters terminate at stable internal contracts, so switching providers does not migrate meetings or retained evidence.
One product, native surfaces
Web, native SwiftUI macOS, iPhone, and Apple Watch share API and design contracts while keeping device-specific capture behavior thin.
What I built
- Bot-free capture across web and native apps, with provider-neutral transcription and conversation intelligence.
- Meeting notes, decisions, and actions that open back to the exact source moments supporting them.
- Automatic deletion of temporary audio and full transcripts after the useful record has been created.
- Responsive web/PWA plus native SwiftUI macOS, iPhone, and Apple Watch surfaces over shared API contracts.
- Encrypted account connections with Gmail as the first integrated workflow.
How it fits the lab
VectorVoice treats memory as a design choice. The system preserves a compact, useful record while deliberately letting the most sensitive raw material disappear.
It turns AI confidence into something a reader can inspect by connecting each important claim to the moment that supports it.
Product and technical scope
I created the product vision and privacy model, designed the information architecture, built the web and native systems, and defined the provider boundaries.