Textrix is the oldest app in this portfolio. It went live in October 2024 as a small iPhone utility for three jobs that all needed a laptop and three browser tabs: paraphrase a paragraph (often into another language), summarise a wall of text, and pull the text out of a PDF. It did them with a cloud model, which in 2024 was the only sensible way to ship.
By 2026 it wasn’t, and the app changed underneath its users. This is the note about why.
Table of contents
Open Table of contents
The original architecture and its problem
Paste text, pick a tool, tap. The app sent the text to a chat-completions API and rendered the answer. Simple, fast to build, and it worked well: the paraphrasing was genuinely good across thirty languages and the summaries were tight.
The problem was the sentence in the privacy section that had to say “the text you paste is sent to a third-party API”. People paste emails, contracts, medical letters and essays into a tool like this. A writing app that keeps a copy of your draft on someone else’s server is a strange thing to sell to people who care about their writing — and a cloud bill scales with exactly the users you’d most like to keep.
The switch
Version 2.8 runs the AI on the device. On phones that support Apple’s on-device foundation model, Textrix uses it — no download, and the text never leaves the phone by construction. On everything else, the app downloads a compact local model (about 940 MB) once and runs it locally. Either way, the tools work with the radio off once set up.
Every feature was re-checked against the local models rather than assumed to work:
- Paraphrase in thirty-plus languages, with eight styles (Standard, Formal, Casual, Academic, Simple, Creative, Expand, Shorten). This is where the quality gap with the cloud model was largest and where the most prompt work went.
- Summarise — the prompt is tuned for load-bearing claims and against bullet-point sludge.
- Grammar — fix spelling, grammar and punctuation while keeping the language, tone and meaning. A proofreader, deliberately not a rewriter.
- Scan to text with on-device OCR, and PDF to text with PDFKit, both of which were always local.
What got worse
I’d rather say this than have a review say it. A 0.6-to-few-billion-parameter model on a phone is not a frontier model in a datacentre. Paraphrases in the long tail of languages are a little flatter than they were. Very long inputs are chunked. The first launch on an unsupported device is a download that the app now warns about — with a specific extra warning on cellular, because nobody wants a surprise gigabyte on a data plan.
What got better
The privacy sentence is now “your text never leaves your iPhone”, and it’s true by construction rather than by policy. Latency on a modern iPhone is comparable to the network round-trip it replaced. The app works on a plane. And the marginal cost of a heavy user dropped to zero, which means the free uses can be generous.
The keyword audit
While the model work was happening I finally tracked Textrix’s App Store keywords properly. After nearly two years live, the app ranked for exactly one term: “proofread”, at position 219. The title had targeted “grammar”, a word with a search popularity of 7 out of 100 — effectively nobody. The tracked keyword list in my tool was polluted with junk from a previous experiment (WhatsApp-in-Arabic, WPS, ZArchiver) that made the dashboard look busier than it was.
The relevant numbers in the US store: “writing” is big and hard; “paraphrase” has decent volume and the best difficulty ratio of anything on the list; “summarizer” is small and cheap; the competitor brand names people actually search for are allowed in the invisible keyword field and nowhere else. So the title became “Writing & Paraphrase - Textrix”, the subtitle “AI Essay & Grammar Checker”, and the eight US-cross-indexed locales were rewritten keyword-first with the brand removed — the Vietnamese listing, for example, carries English tokens like “PDF” and “scanner” that the US index picks up.
Seventeen designed screenshot sets fanned out to all thirty-nine locales, 234 images, uploaded in one fastlane run. And a warning for anyone measuring an ASO change: Textrix had an unexplained organic spike in impressions three days before the new metadata went live. If I’d attributed that bump to the change, I’d have learned the wrong lesson. Lock the baseline first.
Two fastlane gotchas
- A
Deliverfilewithskip_binary_upload truemakes every lane that doesn’t explicitly override it silently skip the binary — then hang “waiting for build”. Took an hour to find. deliverauto-detects the app version from the Xcode project. BumpMARKETING_VERSION(or pass the version in) before the lane runs, or the metadata goes to the wrong version.
What’s next
Textrix is still a subscription (weekly or monthly), which is the one thing in this portfolio I’m not fully at peace with. It predates the “one-time where possible” rule the newer apps follow, and moving an existing subscriber base is its own problem. When the on-device models are good enough that there’s no cloud fallback to fund, the pricing will follow.
Textrix is on the App Store. Every tool has free uses; the model download is honest about its size.