Thought Behind Things
Why teacher grading hours are the wedge for AI in Pakistan
Ansar Moughis, founder and CTO of AIDE AI, explains how the company automates grading of handwritten and digital assessments, why teacher workload was the entry point, and the pricing, trust and expansion problems it faces in Pakistan.
Contents
- Opening: the product in one sentence
- The market question: how many teachers and students?
- How the product works for an ordinary teacher
- Reading handwriting, and where it breaks
- Cheating, detection, and a different kind of assessment
- Pricing and what teachers actually earn
- International schools, government and the local market
- Platform, integrations and the wider product
- Accelerators, and staying rooted in Pakistan
- Human in the loop, and the case for human teachers
Opening: the product in one sentence
Muzamil Hasan opens by placing the episode in an ongoing AI series on Thought Behind Things, which looks at how builders apply AI to different use cases, take products to market, and solve particular problems. He calls the guest’s product “a unique and a very simple use case,” says it has already been validated in Pakistan at roughly one to five thousand dollars in monthly recurring revenue, and notes that the company has been selected for accelerator programmes, with a participant in Riyadh and a co-founder in San Francisco.
Moughis gives the pitch in plain terms: “A teacher spends about 7 to 10 hours every week grading manual papers. We automate that.” The product also gives detailed analytics on students’ strengths and weaknesses so teachers can help them do better. His summary of the aim: making every teacher a better teacher and every student’s success more visible.
The market question: how many teachers and students?
Muzamil presses on market size. Moughis says the company has worked most closely with high schools and has covered K-12 broadly, but the most concrete figure he offers is for O Levels alone: about 15,000 teachers and roughly 400,000 students in Pakistan. He is clear that this excludes FSc and Matric, and that he did not have a full market number to hand. The rough estimate matters more than an exact one, because the total volume of marking hours is large.
How the product works for an ordinary teacher
Moughis describes a web app that also runs on a phone, with no complicated onboarding. A teacher signs in and uploads student papers. The only difference from the existing workflow is that instead of checking by hand, the teacher photographs or scans the pages. The system separates the papers by student, checks them, and returns the results. Muzamil asks whether the phone photos are good enough, and Moughis says they are: “mobiles are good enough now.”
He also describes a route that does not involve teachers at all. Students can open a link, upload their own handwritten work, and close it, with no sign-up. The company has also launched a student-facing practice platform in which ten years of past O Level topical papers are loaded in, so a student can pick a topic and receive a paper matched to its difficulty, then photograph their answers to have them checked.
Reading handwriting, and where it breaks
Muzamil asks how the system reads untidy handwriting, since he understood OCR to handle only printed text. Moughis says the stack is now a combination of LLMs, and that since LLMs arrived, ordinary OCR “got left far behind.” He is candid about edge cases. In one case a student cut out an answer and pasted it on the last page, and the system marked the correct answer when the reviewer expected a wrong one. The team lost a day and a half before realising the last page had been matched. Spelling is another problem: the model tends to correct a student’s misspelling when the test was meant to check it. Mathematics, particularly integrals, is hard, with students producing messy working.
The principle he returns to is that “we never bypass the teacher. There is always a human in the loop.” He says that in about 3% of cases, teachers have changed the answers the system produced. He presents this as evidence that teachers are checking rather than trusting blindly, though he also acknowledges they may simply be checking the system.
Cheating, detection, and a different kind of assessment
Muzamil raises the risk that students copy from chatbots. Moughis agrees the problem is real and that detectors are “all hackable.” He describes a combination of detection, historical handwriting data to flag sudden changes in a child’s style, and a broader point: the assessment itself has to change. Problems should be built so they cannot simply be solved by a model, in the same way that harder questions followed the arrival of the calculator. He argues that “you will have to reimagine how you assess,” with less weight on recall and more on how students reason and put things together.
Pricing and what teachers actually earn
A central thread is cost. Muzamil notes the price is about five rupees per page and argues that the biggest problem for AI use cases in Pakistan is that frontier LLM compute does not make economic sense for local customers. Moughis agrees on the principle and explains that the price was set by working backwards from what schools already pay teachers. He describes most teachers in Pakistan as “grossly underpaid,” and says that teachers who grade large volumes often end up doing it in their own time, so the product is in some cases cheaper than the existing arrangement. He gives an example in which a teacher agreed to about four thousand rupees for a whole set of final papers, saying her week was otherwise consumed.
International schools, government and the local market
Moughis compares the Pakistani market with what he saw during two years in Dubai, where his son attended a local school with a high fee. He says the margins there are wider, so an enterprise customer could fund AI tools for all teachers, while the same margin does not exist in Pakistan. The company works with government on a trial basis, with schools, and with universities, but he notes that large institutions are slow to close, that they tend to test the system and then try to replicate it internally, and that schools in general want a working demo before committing.
Platform, integrations and the wider product
The conversation covers what sits around the grading. Moughis says the company integrates with mainstream platforms such as Google Classroom and offers a public API so that custom LMS providers can connect. Requests to replace an entire LMS come up, but he says the team does not want to be distracted from accuracy in assessment, which is “what we are championing here.” The three things offered to teachers are assignment creation, grading, and detailed analytics. The analytics go down to topic and sub-topic level, showing how a child has performed over the last six months, and there is a chat interface so parents can ask about their child’s progress.
Accelerators, and staying rooted in Pakistan
Muzamil asks about the Riyadh programme and the San Francisco accelerator, Plug and Play. Moughis says the exposure has been valuable, and that being in those rooms helps the company understand the range of what can be built. He is firm that the company still wants to solve for Pakistan, and that it “has to grow outside of Pakistan.” He describes early sales to international schools and a slower, more localised approach in Pakistan, where digital adoption is harder and the team has to sit with teachers and walk them through the product. He also notes that some global competitors have simpler interfaces, and that enterprise buyers abroad will bring different UI, data-security and compliance requirements, which the team is preparing for.
Human in the loop, and the case for human teachers
Toward the end, Muzamil asks whether the way to build for children is to make the system the teacher. Moughis answers that people learn from people, and that isolation makes learning harder, which is why online courses often struggle with completion. He recounts a demonstration to a large school in which the principal said the product would replace all their teachers. His reply was that if the teachers’ marking could be done this way, they should be freed to mentor again. He cites a UNESCO estimate of a shortage of around four million teachers worldwide and concludes that “we will need teachers anyway.” His goal is “every teacher a super-human teacher,” rather than replaced.
Muzamil closes by noting that education is a hard sector to build in because it is a long-term bet, and that the guest is building in it anyway. Moughis says he hopes the company stays in Pakistan and does not “cut off Pakistan” once it has grown, and that solving a local education problem is hard but worthwhile. The episode ends with an invitation for viewers to try the product, share feedback in the comments, and follow Moughis on LinkedIn.
Build with me in the Inner Circle.
You don't need to code. You need a problem you know well, and a room of people building beside you.

