RyzingStar measures how a student actually performs, checks that they are doing the right things, and tells them what to do next. Scores are computed in code. A model writes the advice, and every figure in it is checked against its source before it reaches the student.
Built by a parent for a JEE aspirant at home, who uses it now. It has one user today.
JEE Main is a three-hour paper of 75 questions. A right answer earns 4 marks and a wrong one loses 1. A student rarely loses marks evenly: there are a few chapters he keeps getting wrong, a few minutes sunk into one question, and a few habits under the clock. Generic advice can't see any of these. His own test record can.
So the system starts from the record. It reads every test he takes, works out where the marks go, ranks what to fix by what it is worth in a real paper, and checks that the plan is actually being followed. A coach then answers his questions from that record.
The principle
Correctness over features. The worst thing it can do is show a student a wrong score or a wrong solution. He believes it and acts on it. So there is a deterministic core, agents work on top of it, and every number is checked.
Three pieces, one job each
Measurement
DeepleLens
How did he actually perform?
Marks and negative marking for every test, per subject, in code
A mistake category for every wrong or skipped question
Exam yield: how often each chapter is asked in real papers
The priority board: one ranking of what to work on
Coverage: which chapters his tests have measured, and which they never have
Rank prediction, with its uncertainty shown as a range
A bank of over 13,000 JEE Main past-paper questions (2019–2026) to practise from
An exam-style screen to sit full papers on, timed
Process
Arjuna
Is he doing the right things, at the right time?
The study plan to the exam: phases, months, weekly tasks
Revisions, deep-work hours, sleep and exercise
The Gap Engine: what is being left undone, against targets the family sets
Reads DeepleLens with a read-only key; never writes to it
Action
The coach
So what should he do now?
A Claude-based coach the student and parent reach from a phone
Answers about his data, his plan and the exam, and solves problems
Reads DeepleLens through read-only endpoints
Every answer that advises goes through checks before it is sent (below)
One owner per fact. DeepleLens owns the measurements. The other two read them and never keep their own copy.
How it keeps numbers correct
Scores are computed in code, never generated
Code computes marks, negative marking, the ranking and the rank range, and the same code serves every screen and the coach. Where a model writes a note about a chapter, the note is not allowed to state a figure, and one that does is rejected before it is stored.
The figure gate
Each line of an answer that states a fact about the student names its source, such as a test or a chapter's row on the board. Every number on that line must be found in that source, or the answer is blocked and rewritten. It is a program, not a prompt, so the model can't skip it. Its weak spot is known: a small whole number can match by chance. So a second model also checks the figures by meaning.
Worked solutions are solved twice
Before a solution is sent, a separate solver works the problem from the statement, and only then reads the coach's answer. If they disagree, the coach reworks it. If they still disagree, the answer is sent marked "⚠️ Not verified" and is never presented as certain.
Advice is reviewed before it is sent
A reviewer in a fresh context sees only the question, the draft and the sources. It blocks anything false, unsupported or against the family's rules, and advice that doesn't give its reasons. Every answer that advises carries a "Why" section that can be checked.
Every answer is audited again overnight
An independent run reads the day's questions first and works out its own answers before it sees the coach's, so it isn't anchored by them. A wrong answer is corrected in the chat the next morning. Changes to the coach are proposed, not made by itself, and each fix comes with a test case.
Every answer that advises also comes with its working, as a file: each figure next to the record it came from, the rules applied, and the reviewer's verdict.
Sample data. Everything in this section is made up: the student, the tests and every number. Chapter names come from the public syllabus.
Demo
A made-up student, six chapters. The board is computed in your browser with the same formula the app uses, and the coach's answer is generated from those rows. Then run the figure gate over the answer.
Priority board sample
marks at risk per paper = exposure × failure rate × (1 − practice recovery)
#
Chapter
Asked per paper
Marks lost in tests
At risk per paper
Exposure is questions per paper × 4 marks. The failure rate is pulled toward the student's overall rate by 20 pseudo-marks, so a thin chapter can't top the board on noise. Practice retires at most 60% of the risk. Ranked by marks at risk in a real paper, not by marks lost in tests: notice where Center of Mass lands.
The coach answers sample
Student: What should I work on tonight?
The sources the gate checks against
Status, honestly
Built and in daily use
DeepleLens and Arjuna, live as web apps
The coach, in use from the phone since September 2026
The five checks above, on every answer they apply to
Every coach answer logged, with token use and a thumbs up or down from the family
A set of test cases written from real questions and known failures
Not built yet
The coach reading Arjuna's plan, and writing to it
Splitting the coach into Diagnostic, Planner and Tutor agents with a verifier
An eval suite run in CI, tracing, and cost per run
Anything for a second student: today it serves one, at home
No other users, pilots or revenue yet. The next step is to make it work for other JEE aspirants without losing the checks.
Students, and their data
The users are students, often under 18. Their data stays private:
The student's records are kept in private databases, and the apps are behind a passcode. They never go into a public repository, a demo or this page; everything on this page is made up.
The coach runs under its own restricted account, with read-only access to the student's data. It cannot change the rules it runs under.
A model call sends the records it needs to the model provider: Anthropic for the coach, and one other provider for the chapter notes DeepleLens writes, which are run by hand.
Before anyone else uses it, these rules will be written down for families and agreed with them.
Contact
RyzingStar is built by Nikhil, a solo, bootstrapped founder.