Cuegence

How It Started

The idea began with a very reasonable engineering question: schools collect marks, homework, assignments, teacher feedback, curriculum plans, and years of academic history, so why does almost none of it become a useful model of how a student is actually learning?

This is the sort of question that becomes dangerous when asked by someone who enjoys building systems more than selling them. The correct response is to go for a walk. I opened a whiteboard.

I later wrote the sober version in a Cuegence postmortem. This is what the whiteboard felt like from inside it.

The Pitch

Cuegence was going to be a longitudinal learning-intelligence system for schools.

At the centre was a structured learner model: for every student and every concept, Cuegence would continuously maintain what the evidence suggested they understood, how confident that inference was, what prerequisites might be missing, which misconceptions appeared repeatedly, whether knowledge seemed to be improving or decaying, and which previous interventions had actually helped.

Exam marks, assignments, teacher feedback, and confirmed teaching coverage would map to concepts — evidence of what was taught, not what the annual plan promised.

The important part was that an inference never replaced its evidence. Cuegence would preserve provenance, uncertainty, teacher corrections, later observations, and model history so the whole thing remained auditable instead of becoming an educational horoscope generated by an LLM.

That model could suggest prerequisite refreshers to teachers, send useful practice through channels parents already used, and show leaders where uncertainty remained — without ranking teachers or students.

Student identity would stay outside the intelligence layer where possible, using school-scoped pseudonymous identifiers. Cross-school benchmarking and human rankings were forbidden. Every inference would remain explainable.

By the end, this was no longer a product idea.

It was an architecture diagram with moral principles.

The Delusion

Technically, Cuegence was magnificent: graphs, longitudinal state, evidence provenance, privacy boundaries, and confidence-calibration problems nobody had solved.

The documentation was approaching consciousness. The customer list was not.

Great teachers do form mental models of their students over time. They remember that one child struggles with fractions because multiplication never became automatic. They know which class needs another example before moving on. They notice when a student who normally understands everything suddenly goes quiet.

The market story also looked plausible. Schools already pay for software. AI in education is everywhere. Generic copilots do not know the learning history of a particular class. Cuegence would.

Then came the inevitable founder sentence:

“If we can build the longitudinal learner model properly, this could be a real moat.”

This sentence has funded more architecture diagrams than customer interviews.

Whenever reality objected, architecture offered a pivot: copilot, coaching company, LMS plugin, API.

Every objection had a technical response. Unfortunately, customers are not architecture reviewers.

The Reality Check

I did the deeply unfashionable thing and talked to potential buyers before building it.

Six schools.

Three effectively said they would pay nothing because they did not need it.

Two said something around ₹50,000 — fifty thousand Indian rupees — per year, in the tone of someone donating to a cause.

One said roughly ₹100 per student per month, if Cuegence could replace enough existing school software to justify the cost.

Which was a particularly elegant way of saying, “Your interesting product becomes commercially useful after you turn it into the boring ERP you specifically refused to build.”

Every integration strategy hid the same problem: Cuegence needed continuous, mapped evidence — question-wise marks, assignments, teacher feedback, teaching coverage, and intervention outcomes.

That meant integrating whatever the school already used. The promised one-click plugin was becoming a nine-month install with a WhatsApp group.

The intelligence layer only becomes intelligent after the institution does enough work to feed it intelligence. Until then it is a well-architected empty database with moral principles.

Higher education offered more systems but noisier evidence: increasing sophistication applied to decreasing signal, a category LinkedIn already serves.

Then I pitched Cuegence to my dad.

He has been a teacher for more than forty years.

I explained the learner model. The concept graph. The evidence loop. The Teaching Copilot. The idea that software could continuously understand what each student knows and help the teacher decide what to do next.

His reaction, translated into the polite version, was approximately:

“What is this? I have taught for forty years. I know how to teach children. I know what they know. I know how they will learn. Why would I use software for this?”

There are moments in product discovery when a sophisticated objection requires careful analysis.

This was not one of them.

I had spent weeks designing a system to computationally reconstruct the mental model an experienced teacher already carries around in his head — and my target user was asking why I wanted him to maintain a second, more expensive copy.

Forty years of longitudinal learner modelling, running entirely on tea and repetition, with zero integration cost and full offline support. My competition was my own father, and he had a forty-year head start and no cloud bill.

The problem was real.

The engineering was interesting.

The model might even have worked.

But “interesting engineering problem” and “good business opportunity” had quietly been sitting in different classrooms the entire time.

Why It Stays Unbuilt

Because the strongest version of the product required schools and teachers to feed a complex intelligence system so it could help them do something many of them already believe they know how to do.

Six schools gave me pricing data. My dad gave me the answer.

Customer discovery: one sentence. Time to insight: forty years, but they were his, not mine.

Unwanted Bonus

Cuegence Teacher Pro.

Every evening, the teacher spends ten minutes entering observations about the class so the AI can analyze them overnight and send a beautifully formatted notification the next morning:

“Based on 47 longitudinal signals, you may want to revise fractions today.”

The teacher looks at the notification.

The teacher had already planned to revise fractions today.

The teacher enters that as feedback so the model can improve.

The model files it under evidence provenance.

Confidence: increasing. Usefulness: longitudinal.

---

I thought about building this and chose not to.