I built an iOS app where AI researches your shower thoughts (and what it actually costs)

mdl2170 — Tue, 21 Jul 2026 14:04:12 +0000

I kept losing ideas. Not for lack of capture tools — I had Apple Notes, Notion, and a graveyard of half-sentences like "app for the thing??" that meant nothing a week later. The problem was never capture. It's that nothing happens after capture.

So I built Thought Spark: a solo-built iOS app where the loop is the product — Capture → Clarify → Research → Act. Here's the architecture, the surprises, and the real unit economics of shipping an LLM app alone.

The stack

iOS: SwiftUI + SwiftData, on-device-first. Ideas never live on my server.
Backend: a deliberately thin, stateless FastAPI service on Render. It holds the API keys, meters usage per user, and does nothing else. The server stores identity + usage counts — not idea content.
AI: Claude (Opus for reasoning/research, Haiku for cheap cleanup), ElevenLabs Scribe for transcription.
Auth/billing: Sign in with Apple → backend-issued JWT; StoreKit 2 with server-side receipt verification via the App Store Server API.

Surprise #1: on-device speech recognition can't do multi-sentence

I started with Apple's Speech framework (requiresOnDeviceRecognition = true) for privacy. It works beautifully — for one sentence. The recognizer resets its formattedString per utterance, and there's no clean signal to commit the previous one. After days of segment-stitching hacks, I switched to recording an m4a and uploading to ElevenLabs Scribe (~$0.40/audio-hour). Then a small model pass cleans the transcript: "Um so like I was thinking…" becomes prose. That cleanup step is the single most-loved thing in the app and it costs ~$0.003 per capture.

Surprise #2: research UX is a progress problem, not a prompt problem

The research feature sends an idea to the web (Claude's web-search tool) and returns prior art, gaps, and sources. v1 was one big call behind a spinner — 60–90 seconds of blind waiting. Terrible.

The fix: a /research/plan endpoint first returns a 3–4 step plan the user can edit, then each step executes as its own call with persisted state. Progress is real ("Researching 2 of 4"), results stream in as steps finish, failed steps retry individually, and the run survives app navigation because state lives in SwiftData, not in a view model.

The economics (real numbers)

Operation	Cost/call
Capture (transcribe + clean)	~$0.003
Title/tags	~$0.003
Clarify interview turn	~$0.01–0.02
Web research run	~$0.30–0.50

Research dominates — one run costs more than a hundred captures. That's why the free tier is generous on capture (unlimited, forever) and capped on research (3/month), and Pro ($7.99/mo) has a 50-run fair-use cap. The pricing isn't growth hacking; it's the cost structure made legible.

What I'd tell past me

Meter from day one. Unauthenticated LLM endpoints on a public URL are a wallet with a URL.
Don't fight platform speech recognition; it's not built for dictation-length input.
Progress UI beats faster inference. Users forgive 90 seconds they can watch.
StoreKit 2 + App Store Server Notifications V2 is genuinely good now — but verify signatures server-side; never trust the client's "I'm Pro."