Skip to content
TP
← All work
LiveApr 2025 – Mar 2026

ThirdSlate

A study platform that reads the actual course, not the internet.

Next.jsTypeScriptPythonFastAPIPostgreSQLpgvectorAWS LambdaStripe
thirdslate-prase.app

PHYS 216 · unit 4

7 / 12

Which quantity is conserved when the net external torque is zero?

source · lecture 11, slide 6

ALinear momentum
BAngular momentum
CKinetic energy
DMechanical work

verified against source

Course-scoped

Generation is bounded to the unit being revised, not the whole uploaded library.

Cited source

Every question carries the lecture and slide it came from, which is what makes it trustworthy.

Verified answer

The answer is checked against its source passage before a student ever sees it.

Hover a marker to see what each part of the interface does.

Students served

01

1,000+

cumulative registered users

Questions generated

02

{{QUESTIONS_GENERATED}}

Retention

03

{{RETENTION_METRIC}}

week 4 return rate

Cost per study session

04

{{COST_PER_SESSION}}

01

The problem

What was broken before any of this existed.

Generic AI study tools generate plausible questions that miss the course. Students revise the wrong material, get a good score in the app, and get surprised by the exam.

University students revising for midterms and finals, mostly on a phone, mostly the night before.

Constraints I worked inside

C01Two founders, no budget for inference waste
C02Course material arrives as messy PDFs and slide decks, not clean text
C03Trust dies on the first wrong answer, so precision mattered more than volume
90

ownership

end to end

100

role span

of four tracks

85

shipped

reached users

02

What changed

The same job, before and after the work.

Before

  • Generic tools generated plausible but off-syllabus questions
  • Students revised the wrong material
  • No link between a question and the course
  • Trust lost on the first wrong answer

After

  • Every question scoped to the unit being studied
  • Answers verified against the source before display
  • Citations back to lecture and slide
  • 1,000+ students, no paid acquisition
03

The system, end to end

How a request travels through it.

INGESTcourse PDFs, slides an01EXTRACTadaptive structured ex02RETRIEVEpedagogically scoped r03GENERATEquestions written agai04VERIFYanswer checked back ag05
01

course PDFs, slides and notes parsed into structured units

02

adaptive structured extraction keyed to learning objectives

03

pedagogically scoped retrieval, bounded to the unit being studied

04

questions written against retrieved course content with citations

05

answer checked back against the source before it reaches a student

04

Where the effort went

Effort split and how quality moved.

$stat --effort

Where the time actually went

effort
  • Ingestion and extraction42%
  • Retrieval and generation33%
  • Product, payments, growth25%

$plot --trend

How it moved

sepoctnovdecjanfebmarapr
cumulative student signups, no paid acquisition
05

Decisions

Every call I made, and the tradeoff I took for it.

  1. Why

    A question drawn from week 11 during week 3 revision is technically correct and practically useless.

    Tradeoff accepted

    Fewer questions available per session, and worse behaviour when a student uploads badly organised material.

    Weighting

    gain
    cost

    my judgement, not a measurement

  2. Why

    One confidently wrong answer costs more trust than ten slow ones.

    Tradeoff accepted

    Roughly double the inference cost per question, which forced hard work on caching and batching.

    Weighting

    gain
    cost

    my judgement, not a measurement

  3. Why

    I needed to know which behaviour to charge for. Guessing at pricing before that is expensive theatre.

    Tradeoff accepted

    Months of unmonetised inference spend, absorbed deliberately as the cost of learning what people used.

    Weighting

    gain
    cost

    my judgement, not a measurement

  4. Why

    Session analytics were unambiguous about when and where students actually revise.

    Tradeoff accepted

    The desktop experience stayed plainer than I wanted for longer than I wanted.

    Weighting

    gain
    cost

    my judgement, not a measurement

06

Outcome

What was true after it shipped.

01

Grew past 1,000 students with no paid acquisition, mostly through course-level word of mouth.

02

PRASE became the differentiator competitors could not match with prompt engineering alone.

03

Set up the pricing and packaging decisions with real usage data instead of assumptions.

What worked

Treating trust as the core metric. Verification felt expensive on the spreadsheet and turned out to be the reason students came back.

What was hard

Ingestion. Real course material is far messier than any demo dataset, and most of the engineering time went there.

What I would change

I would instrument the learning outcome, not just engagement. Time in app is a weak proxy for whether someone actually did better.

$./next-step

Hiring for a product engineer who can carry the whole thing?

Send me the role and the problem behind it. I will tell you honestly whether I am the right fit.