Skip to content
Sid Prithvi
← Index
02 · Mobile IELTS writing coach · Technical co-founder · 2025 · Built

UpBand

UpBand is an IELTS Writing coach for phones. It scores an essay against the four criteria a real examiner uses, tells you which sentence to fix, and turns vocabulary into a daily habit rather than a list you reread. There are two of us on it. Product and engineering are mine, so everything from the app down to the database is something I wrote.

Abstract mark: an essay beside a band scale, with the band it landed on

The Problem

IELTS Writing is scored on four criteria: Task Achievement, Coherence and Cohesion, Lexical Resource, and Grammatical Range and Accuracy. Each one runs 0 to 9. It's the section people most often underperform in, and it's the hardest one to study on your own.

There are three walls if you're preparing alone. You get no feedback signal, so you can write twenty essays and have no idea whether you're improving, and having one marked by a human costs around $50. You hit a vocabulary ceiling, and Lexical Resource is a quarter of the grade. And you have no curriculum, because “practise more” isn't a study plan.

The Decisions

  • 01

    The grader shows its reasoning

    It doesn't just return a band score. It says which descriptor it applied and why, quotes the sentences that cost you marks, and rewrites them. A number on its own teaches you nothing, and nobody trusts a number they can't argue with. The quoted line and the rewrite are the part that turns feedback into something you can act on tomorrow morning.

  • 02

    The grading prompt never ships inside the app

    Evaluation runs in Supabase Edge Functions rather than on the phone. Keeping the API key off the client is the obvious reason and the smaller one. The real reason is that the prompt encodes the rubric, and the rubric is the product. Ship it in a binary and anyone who wants it can unpack it.

  • 03

    Half the drills never call a model

    Sequencing and question-type identification have deterministic right answers, so rules check them. Only paraphrasing and grammatical structures go to the model. Sending a question that has exactly one correct answer to an LLM spends tokens and adds a round trip to buy nothing. It's the less impressive answer on a slide. It's also the difference between a demo and something I can afford to leave running.

  • 04

    A wrong answer costs you one level, not all of them

    Cards sit on a six-level ladder, with reviews at 0, 1, 3, 7, 14 and 30 days. Get one right and it moves up a level. Get it wrong and it drops one, instead of resetting to zero. Resetting is the usual behaviour and it's the crueller one. A bad morning on a word you've known for weeks shouldn't wipe a month of scheduling. What it costs me is that a genuinely forgotten word takes longer to work its way back down to daily review. I'd rather that than have someone stop opening the app.

  • 05

    A dead network gives you a worse score, not an error screen

    When the evaluator can't be reached, the drill screens fall back to a local heuristic check. You still get a score, a model answer, and credit for the session. It's less accurate than the real thing. It's also much better than dead-ending someone who sat down to study. The session stays useful when the network isn't.

What Shipped

Three surfaces, tied together by one progress system. Encouraging, never childish.

  • Cards

    452 academic phrases in 9 categories. Overdue cards come first.

  • Skills

    Writing split into 4 drills, one sub-skill at a time.

  • Practice

    The full exam. 131 prompts and 48 charts. Plan, check the plan, then write.

  • Locked down underneath

    63 row-level security policies. The AI box is treated as hostile input.

  • React Native
  • Expo
  • TypeScript
  • NativeWind
  • Supabase
  • Postgres
  • Claude
  • Maestro
  • React Native
  • Expo
  • TypeScript
  • NativeWind
  • Supabase
  • Postgres
  • Claude
  • Maestro
Home screen: today's goals and weekly progress, with a card deck to pick up where you left off
A vocabulary deck: 85 expressions, how many are studied and mastered, and how many are due for review
A flashcard mid-session, tagged intermediate, with tap to flip and swipe to answer
A Task 2 essay question with planning tips before you start writing
The question list for one essay type, with the ones already completed marked
Paraphrase feedback: eight out of ten, with the reason and a meaning-preserved check
Task 2 essay types, each with a count of how many are done

Where It Is Now

Built end to end. Two of us on the company, with product and engineering mine. The scope is real rather than a prototype's: about 60 screens, 26 migrations, two Edge Functions, and content tables with 452 phrases sitting in them.

What I can't show you yet is whether it moved anyone's band score. That's the only number that would prove the thing works, and getting it means putting it in front of a cohort and waiting for them to sit the exam.