Sathvik VK

Flow

A Mac app that turns speech into text in any field, drafts replies from what is on screen, and keeps its memory on the machine.

Status
Beta. Private beta, shipped as signed and notarized builds. Version 0.12.0 in Nov 2026
My role
Lead engineer. I wrote most of the app, with a small team contributing.
When
Oct 2026 to present
Context
Perfstaq product
Capabilities
LLM apps, Agents, Memory, Voice, Desktop, Integrations
Three frames of the Quinn panel: a request to send an NDA, a plan with checked steps, and an approval card before sending.
Quinn handling a request: plan, approval card, result. Frames from the app's own UI recording, with sample data.Screenshot

What it is

Flow started as dictation: hold fn, speak, and the text lands in whichever app has focus. Speech recognition runs locally through a bundled whisper.cpp, so audio does not leave the Mac.

It grew two more parts. A writer drafts text for the field you are in, using what is on screen and what the app remembers about the person you are writing to. Quinn, added in 0.12.0, is an assistant that answers from the same memory and can prepare actions such as sending an agreement. Quinn and Flow are one app and one codebase, so I count them as one project.

What I built

I wrote the Electron main process and most of the features on top of it, and the Swift helper that does what Electron cannot: global key handling, reading the focused text field, inserting text, screen capture and OCR.

  • Dictation pipeline: key hold, 16 kHz capture, local transcription, insertion into the focused field.
  • Writer flow in four steps: locate the field, build a brief, review, insert.
  • A local memory in SQLite with full-text search, fed by on-screen activity and read by both the writer and Quinn.
  • An MCP server inside the app, so other AI tools on the same Mac can query that memory. It is connected to Claude on my own machine.
  • Release pipeline: Developer ID signing, notarization, and builds published from CI.

Decisions

The writer prefers Apple's on-device models and falls back to a local Qwen3 through Ollama, so the default path needs no cloud model. Sending memory to the cloud is off unless the user turns it on.

Quinn follows one rule I use across projects: the model proposes, code checks, and a person approves. Anything that changes the outside world is put on a card the user has to hold to confirm.

When the app's display name changed to Perfstaq, I kept the bundle identifier and data directory. Renaming them would have reset macOS permissions and broken every existing MCP configuration.

How it fits together

Input
Swift helper
keys, focused field, screen capture, OCR
Microphone
16 kHz audio
On the Mac
whisper.cpp
local speech to text
Memory
SQLite, full-text search
Writer and Quinn
on-device model, Ollama fallback
Check
Approval card
hold to confirm any action
Output
Text in any app
inserted by the helper
MCP server
memory for other AI tools
Speech and screen context go in at the top; text and approved actions come out at the bottom. Simplified diagram drawn for this page, not a product screenshot.

Built with

ElectronSwiftwhisper.cppSQLite full-text searchApple on-device modelsOllamaQwen3MCPPlaywright
Next case studyOr press the right arrow keyPerfstaqThe company I founded. A workspace of business apps (Sign, Board, Sales, Calendar, Notes, Finance) that AI assistants can operate through one connector, with approvals.