The grading was never the teaching.
Marking, notes, admin — the work that fills a teacher's week is the work that least needs a human in it. Stargaze automates that layer, so the hours left over go where only a person can go.
Each tag is written to the student's error profile.
A teacher's scarcest resource is attention.
Thirty scripts take a weekend. Notes take an evening. Chasing who submitted what takes the rest. None of it is teaching — it is the toll paid before you are allowed to teach, and it is paid in the hours you would rather spend noticing that a student has quietly given up.
Stargaze inverts the order. The machine takes the marking, the drilling, the notes and the admin. What it hands back is not a grade — it is a teacher's week.
I didn't build this to replace teachers. I built it so I could afford to be one.
Aaron Li · Founder
A system cannot make a student work. It can clear everything standing between a teacher and the moment a student needs them — and it can make sure that when the work does come in, the feedback is already waiting, specific, and aimed at what that student in particular got wrong.
Four papers. One system.
The HKDSE English examination is four papers, each with its own rubric and its own kind of examiner. Stargaze is building an engine for every one of them, on shared infrastructure.
Comprehension analysis and paper construction from a structured question blueprint.
Full diagnostic marking, corrected script, targeted drills and a rewritten exemplar. Serving students today.
Page-by-page automated marking of scanned answer books, with examiner override.
Transcription and group-discussion assessment, adjudicated by more than one examiner pass.
Assessment runs as a panel, not a prompt.
A single model asked to “grade this essay” returns a plausible number and little else. Stargaze decomposes the judgement the way a marking panel does, so every score arrives with a reason a teacher can defend to a parent.
Intake
Handwritten scripts are photographed; typed work is submitted directly. Both normalise to the same structured text.
Parallel diagnosis
Several specialised passes read the script independently and at the same time, each producing evidence rather than an opinion.
Adjudication
A senior examiner stage reconciles that evidence into rubric scores and writes the justification for each. Disagreement is resolved here, not averaged away.
Composition
The report is built to fit the script it came from: corrections, drills aimed at the errors that actually appeared, and a rewritten exemplar pitched one band above where the student is now.
A mistake isn't corrected until it stops being made.
The report is not the end of the transaction. Everything the engine finds is written to a per-student error profile, and that profile decides what the student meets next.
Errors become cards. Vocabulary and structures the student got wrong enter a spaced-repetition deck automatically, with British pronunciation rendered server-side so every student hears the same voice.
Cards become quizzes. Timed and marked on the server, drawn from that student's own history. Answers never reach the browser.
Quizzes become practice. Scaffolded exercises generated from the specific rules a student keeps breaking, not from a generic bank.
Practice becomes the next script. The next submission is measured against the same profile, so progress is a curve rather than an anecdote.
One student, on one screen.
Every script, quiz and exercise a student completes lands in the same profile. A teacher opens one page and knows what to do next — without reading thirty scripts to find out.
Writing score · last eight submissions
Latest script · rubric breakdown
Recurring errors · driving this week's revision
Built to be handed to other teachers.
The platform was written by a working tutor for his own students, then generalised. An institution runs on its own tenant, with its own teachers, cohorts and materials.
Your teachers, your cohorts
Teacher accounts are provisioned per institution; student cohorts import in bulk or join by invitation.
Your papers, from your notes
Author a paper from your own teaching material. It reaches students in the app and prints as a watermarked A4 script.
The teacher holds the release
Machine marks are held for approval. Nothing reaches a student until a person releases it.
Cohort-level diagnosis
The profile that drives one student's revision aggregates into what a whole class is getting wrong.
Built end-to-end on Google Cloud.
The assessment engine is not a feature bolted onto a website — it is the company's core infrastructure. Marking is dispatched as asynchronous work, so an assessment that takes minutes never depends on a browser staying open.
Aaron Li
Aaron Li founded Stargaze Education in 2025. He teaches HKDSE English in Hong Kong and is an undergraduate at the Chinese University of Hong Kong.
The platform began as internal tooling. He was marking every script by hand, writing every set of notes, and building every mock paper — including two full public mock examinations, all four papers with original listening audio. Automating that work was not a product decision. It was the only way to keep teaching at the standard he wanted while the number of students grew.
He returns each year to run his secondary school's elite English class, and in 2026 partnered with HKFYG Farm Road Youth S.P.O.T. to deliver free DSE preparation — workshops, mock examinations, speaking practice and essay marking — to the community. He has sat the DSE English paper himself, most recently at 5**.
Teaching is not a side effect of the company — it is how the system is tested. Every capability in the platform exists because a real script exposed the need for it.