TIM SCHROEDER
Work / Infinity Dictate
Voice AI · macOS app

InfinityDictate

A Mac dictation app that turns speech into clean, ready-to-send text in any app. Speech is transcribed on the Mac by default, and AI rewrites it on request.

Live · on holdFeb–Jul 2026
Infinity Dictate homepage with the headline 'Stop Typing. Start Thinking.' above a mock recording windowReal screenshot · public homepage

The problem

Typing is slower than speaking. The Mac's built-in dictation gives raw text that still needs fixing before it can be sent.

Who it's for

Mac users who write a lot: email, notes, documents and messages.

Where AI is used

An on-device Whisper model turns speech into text. A language model then cleans up the text or rewrites it into one of 27 formats, such as an email or meeting notes.

Built with

  • Swift 6 / SwiftUI
  • WhisperKit
  • Supabase
  • OpenAI
  • Claude
  • Stripe
  • Next.js
  • Sparkle
01 · The problem

Speaking is fast. Fixing the text is not.

Most people speak much faster than they type. But dictated text arrives with filler words, missing punctuation and no structure. Someone still has to tidy it up.

I wanted one hotkey that works in any app. You speak, clean text appears where the cursor is, and you can turn it into an email or a list without copying and pasting.

  • Raw dictation needs manual cleanup before it is usable.
  • Turning notes into an email or a summary means a second tool and copy-paste.
  • Sending your voice to a cloud service is a privacy worry for many people.
  • Dictation has to work in every app, not only in one editor.

Why custom software

No single tool I had did all four things at once: on-device speech, AI rewriting, typing into any app, and its own plans and usage limits. So I had a native Mac app built.

02 · The solution

A menu-bar app that types clean text for you.

Press a hotkey and speak. The app transcribes on the Mac, applies your voice commands and dictionary, and types the result into the app you are using.

AI

On-device transcription

WhisperKit runs a Whisper large-v3-turbo model on the Mac and offers 99 languages. This is the default, so speech stays on the computer unless a Pro user opts into the cloud engine.

AI

AI cleanup and 27 rewrite formats

A language model fixes the text or rewrites it as an email, bullet points, meeting notes, a commit message and more. Emails, phone numbers and card numbers are masked before text goes to a cloud model.

AI

Live typing and voice commands

Words appear as you speak. Commands like “new paragraph” and “scratch that” edit the text as you go.

Works in any app

Text is typed through macOS Accessibility, with a clipboard fallback for apps that refuse it. Push-to-talk and a toggle hotkey are both supported.

AI

File transcription

Audio and video files are transcribed on the Mac with the same Whisper model.

Plans, quotas and an admin console

Stripe checkout, a server-side entitlement engine and a Next.js admin console with 25 API routes for usage, billing, alerts and users.

03 · My role

What I did. What the AI did.

Claude Code wrote all of the code, including every line of Swift. I decided what the app should do, which models and services it uses, what to test, and when it was ready to ship.

Tim · human judgement

I owned

  • Product concept and scope: a system-wide Mac dictation app with AI rewriting.
  • Privacy direction: keep transcription on the Mac by default. In March I dropped the Deepgram cloud engine and moved all speech to on-device WhisperKit.
  • Model choices, based on a written cost comparison of five models: GPT-4.1 mini as the default, Claude Haiku 4.5 as the failover.
  • Plan design: a free tier, a Pro tier and a 7-day Pro trial.
  • Commissioned the audits: security, red team, UX and a full-codebase bug review. I chose which findings to fix first.
  • Release: Apple Developer ID signing, notarization, the Sparkle update feed and the public website.
Claude Code · AI implementation

AI did

  • Wrote the Swift app, the 13 Supabase edge functions, the database migrations and the Next.js admin console.
  • Ran the audits as parallel review agents, with a second agent trying to disprove each serious finding.
  • Diagnosed bugs from my reports and the app logs, and drafted the fixes.
  • Wrote the unit tests, the edge-function tests and the build and release scripts.
  • Ran compile checks (swift build, tsc, deno check) before each change landed.

I cannot write Swift myself. My job was to describe the behaviour, test it in real use, and push back when the result was wrong.

04 · How I built it with AI

From idea to production.

The project ran from February to July 2026: 253 commits, 240 of them with a Claude co-author line. It started from a starter kit for AI coding agents called “Anti-Gravity”, which is why the README is still the kit's own text.

  1. 19 Feb 2026 · Build

    First commit: a working dictation pipeline

    The first commit already had speech-to-text, AI refinement, text injection, live typing, history and Supabase sign-in.

    commit 0f5dc50
  2. 20–23 Feb 2026 · Harden

    Security audit, billing and an admin console

    A security audit and a red-team pass came first. Then Stripe, a server-side entitlement engine and the admin console. A cost comparison set GPT-4.1 mini as the default model.

    commits e2c8a7d, 20c6b4a · LLM_Model_Cost_Analysis.txt
  3. 3 Mar 2026 · Pivot

    All speech moves onto the Mac

    I removed the Deepgram cloud engine and its two edge functions. Every user now transcribes on-device with WhisperKit.

    commit 80ac800
  4. 4–6 Apr 2026 · Release

    Version 2.0.1 ships as a notarized download

    A signed, notarized and stapled DMG with a Sparkle update feed. A second build variant was added for the sandboxed Mac App Store.

    commits e75672f, 0a30aa6
  5. 8–12 Apr 2026 · Fix

    Live typing stops duplicating text

    Six fixes in three days for live typing that repeated text on screen. An opt-in cloud speech engine was also added for Pro users.

    commits f7277f1 → b765d1f, f334c44
  6. 10 Jun 2026 · Review

    A full-codebase bug review: 46 bugs fixed

    Six review agents read the app, the edge functions and the admin console. 1 critical, 15 high and 30 medium bugs were confirmed and fixed, followed by about 24 low ones.

    commits b6b23e3, a2406e5 · CODE_REVIEW_FINDINGS_2026-06-10.md
  7. 25 Jul 2026 · Website

    Last change: a faster homepage

    Render-blocking fonts and icons were removed. PageSpeed mobile went from 65 to 75 and desktop from 93 to 99. The project has been on hold since.

    commit dc89067 · memory/log.md
05 · Technologies

What it runs on, and why.

Every item here is in the project's code or configuration. I chose and connected them; AI coding agents wrote the code that uses them.

AI models

WhisperKit (Whisper large-v3-turbo)
On-device speech-to-text, the default for every user. Apple Speech covers the gap while the model loads.
OpenAI gpt-4o-mini-transcribe
An opt-in cloud speech engine for Pro users, routed through a server proxy.
OpenAI GPT-4.1 mini
The default model for cleanup and rewriting. Chosen after a per-call cost comparison.
Claude Haiku 4.5
Failover when OpenAI is unreachable, and the model for new users' trial calls.
Ollama
An optional local model, so rewriting can also run without the cloud.

Mac app

Swift 6 + SwiftUI
The native menu-bar app, built with Swift Package Manager. Targets macOS 14 and later.
Accessibility API
Types text into other apps, with a clipboard fallback.
Sparkle
Signed in-app updates for the direct download. Left out of the App Store build.
PostHog
Product analytics events.

Backend

Supabase
Sign-in, Postgres with row-level security, and 13 Deno edge functions.
Edge proxies
llm-proxy and transcribe-proxy hold the provider keys and check the user's plan and limits on every call.
Stripe
Checkout and webhooks, with signature checks and duplicate-event protection.

Ops

Next.js 16 admin console
Usage, costs, billing, alerts, feature flags and user management.
GitHub Actions
Builds the Swift app and type-checks every edge function on each pull request.
Developer ID + notarytool
A release script signs with hardened runtime, notarizes and staples the DMG.
Vercel
Hosts the public website.
06 · System architecture

How the pieces fit.

Speech is turned into text on the Mac. Only text, or audio from users who opt in, goes to the cloud, and always through Infinity Dictate's own server first.

Infinity Dictate architectureSpeech is turned into text on the Mac. Only text, or audio from users who opt in, goes to the cloud, and always through Infinity Dictate's own server first.audiorequeststextfailoverquotasreadswebhooksMac appSwift 6 · SwiftUIOn-device speechWhisperKit · WhisperEdge functionsSupabase · Deno · 13PostgresSupabase · RLSAdmin consoleNext.js 16OpenAIgpt-4.1-mini · transcribeAnthropicClaude Haiku 4.5Stripecheckout · webhooks
InterfaceApp / serverAI serviceData / storageExternal service
Technical detail for engineers
  • In a release build the app never calls a cloud model provider directly. llm-proxy and transcribe-proxy check the user's sign-in, plan, feature flags and daily caps first.
  • llm-proxy only accepts models on an allowlist and caps prompt size. If OpenAI fails, it retries on Claude Haiku 4.5.
  • Before text leaves the Mac, card numbers, phone numbers and emails are swapped for placeholders. They are put back when the result returns.
  • The entitlement engine issues signed tokens with an offline grace period. Rate limiting fails closed.
  • A compile flag builds two variants from one codebase: the direct download (Accessibility typing, Sparkle) and a sandboxed App Store build (clipboard typing only).
  • One of the 13 edge functions, vision-proxy, is retired. It now returns 410 Gone.
07 · Challenges

Problems worth telling you about.

Real incidents from the project's history, each with how it was found, fixed and checked.

CHALLENGE 01

Live typing typed the same words four or five times

With live typing on, text sometimes repeated on screen. Long sentences and voice commands like “new paragraph” made it worse.

VERIFIEDAfter the final fix, the next commit moved on to other work. Live typing has no automated test; the April regression plan lists it as manual-only. The June review later found and fixed two rarer duplicate-text paths.
Why it was hard, what I decided, how AI helped
Why it was hard
Whisper keeps revising its guess for the whole sentence while you talk. The app has to work out which characters are already on screen and change only the new ones. The earlier fixes each patched one case and missed another.
My approach
I kept reporting each case I hit, and pushed for one root cause instead of more patches.
How AI helped
Claude Code found that the app compared raw words with text where voice commands had already been replaced. The match failed, so the whole transcript was typed again. The fix tracks only the most recent window of text, raw against raw.
The decision
Replace the logic instead of adding a sixth patch. The final fix removed about 120 lines of matching code.
Evidence
commits f7277f1, 8a1b772, 4315586, ccbcde7, 57fbc2a, b765d1f · REGRESSION_TEST_RESULTS.md
CHALLENGE 02

A critical bug in shipped file transcription

Most audio files are recorded at 44.1 or 48 kHz. The app passed them to Whisper as if they were 16 kHz. Whisper heard them about three times too slow, so the text was poor, and logged usage was about three times too high.

VERIFIEDThe finding was confirmed against WhisperKit's source code and the build compiled cleanly after the fix. No automated test with a real 44.1 kHz file was added.
Why it was hard, what I decided, how AI helped
Why it was hard
Nothing crashed and the build was clean. Live dictation was fine because the microphone path already converted to 16 kHz. Only one WhisperKit entry point skips resampling.
My approach
I commissioned a full-codebase review with six parallel agents. A second agent had to try to disprove every critical and high finding before it counted.
How AI helped
The review traced the bug to WhisperKit's own source, where the sample rate is fixed at 16,000. Claude Code then added a conversion to 16 kHz mono and corrected the duration maths.
The decision
Fix all 46 confirmed bugs in one pass, then the low-severity ones, instead of shipping only the critical fix.
Evidence
commit b6b23e3 · CODE_REVIEW_FINDINGS_2026-06-10.md (C1)
08 · Results

What is true today.

Infinity Dictate was built and released as a signed Mac download. It has no paying customers and has not been in active development since July 2026, so these results describe what was built.

253
commits, Feb–Jul 2026
Source: GitHub commit history, 2 Nov 2026
31.6k
lines of Swift in 85 files
Source: git ls-files *.swift | wc -l
13
Supabase edge functions
Source: supabase/functions/
46
confirmed bugs fixed in one review pass
Source: commit b6b23e3
27
AI rewrite formats
Source: RewriteFormat.swift
99
languages for on-device transcription
Source: WhisperEngine.swift
  • Version 2.0.1 shipped as a Developer ID–signed, notarized DMG. The DMG passes Apple's stapler check.
  • infinitydictate.com is live, with a Sparkle update feed for the app.
  • On every pull request, CI builds the Swift app and type-checks all edge functions.
  • The repo has 63 Swift unit tests and 35 edge-function tests. That is thin coverage for the size of the app.
  • A sandboxed Mac App Store build exists, but the App Store submission was never completed.
09 · Lessons

What this project taught me.

  1. Reading code is not the same as testing it. In April a check marked the Stripe failure path as passing by reading the code. In June the review found it still told Stripe not to retry.
  2. Test with real inputs. The critical audio bug shipped because nothing fed the app an ordinary 44.1 kHz file.
  3. When a fix needs a fifth attempt, simplify the design. The duplication bug went away when the tracking logic got smaller.
  4. Price the model before you choose it. A written cost comparison made GPT-4.1 mini the default and Claude Haiku the failover.
10 · Demo

See the product site.

The website is live and shows the app and its features. The Mac app is not in active development, so there is no guided tour of it here.

Let's talk about
what you need built.

Open to AI roles: full-time, contract or consulting. I build AI products, internal tools and automations with AI coding agents.