Meet QuestionWell: a K-12 platform that uses AI to help teachers elevate the quality of supplemental curriculum, generating standards-aligned assignments and assessments in minutes. The platform exports directly to 20+ classroom tools teachers already use, including Canvas, Kahoot, and Google Classroom.
Beyond assessments, QuestionWell has expanded into a broader suite of AI-driven teacher workflows that reduce lesson planning time while elevating the pedagogical quality of teaching materials. They do this by developing research-aligned agents and tools that support teachers in seeing the aspects of pedagogical quality they find important, applying those aspects of quality to their own curriculum as they localize it to their students, and then using AI to iteratively refine materials. All content generation and analysis is powered by QuestionWell’s own fine-tuned AI models, purpose-designed and validated for educational material rather than general-purpose AI.
Our conversation with co-founders Maya Bialik and Will Cummings explores how a classroom problem became a question-generation engine, which then became an entire platform for informed decision making around learning experience design, and why they believe instructional quality has to be both visible and actionable for teachers.
You founded QuestionWell with a research and technical background. What was the gap you found in the classroom that made you think “We have to build this”?
Maya: I was a middle school science teacher at BC High, and the problem I kept running into was that assessment set the tone for everything. Even in seventh grade, even in the middle of a fun demo, students were focused on the test, asking how many retakes they’d get, weeks before I’d even mentioned an assessment. I could feel the anxiety building around it, and I wanted to offer them more: more retakes, and more practice opportunities in general. My background is in cognitive science, so I knew how important practice is for learning, but before AI, I simply didn’t have the time to write that many retakes. I tried using ChatGPT to help, but it was bad at writing multiple-choice questions; it kept running into the same handful of problems. Around that time, Will and I happened to be working right next to each other. I was building a quiz on atomic structure, and he was experimenting with a trivia question generator during a Hack Week at Spotify. We realized we were circling the same problem from two different directions, and that AI could with the right prompts and tuning probably write much better questions than either of us had assumed.
Will: I did a lot of work with Maya on this early on, starting with fairly basic prompt engineering and then moving into more sophisticated approaches as we learned more. It helped enormously that we had a rapid iteration cycle, since Maya was actively teaching with it in real time. She’d come back with very specific, practical feedback, things like all the answer choices in a question being correct, or the wrong answers being too close to the right one to clearly distinguish.
What’s the biggest difference teachers notice once they start using QuestionWell?
Maya: It depends on what they were doing before. Teachers writing questions by hand notice the speed. Teachers already using other AI tools notice the quality.
Will: At the most basic level, we’ve built our own platform for developing and evaluating prompts and fine tuned models, so we can make changes that actually improve instructional quality for the teachers, not just staying flat or getting worse. Early on, that meant having a tight feedback loop with teachers actively using QuestionWell in the classroom. We’d get specific notes, like when distractors were too close or an answer key was wrong, and use that to refine our prompts. That groundwork, done with earlier and more limited models, means we benefit even more as the underlying models improve, which keeps our margins strong.
Maya: QuestionWell has also grown well beyond generating multiple-choice questions. It now adjusts materials for the student to specific reading levels, generates lesson plans, and can analyze them according to research constructs, offer suggestions for improvement, and refine them too. In every case, the goal is the same: make sure the output is pedagogically sound, and equip teachers with the tools to adjust materials for their students.
Will: Beyond that, we surface pedagogical quality markers that can be hard to judge at a glance, like how a lesson plan scaffolds cognitive load, or whether it maintains a gradual release of responsibility as more scaffolding is added. That lets teachers refine outputs in an informed way, instead of iterating with AI blind.
Maya, you’ve been a teacher and researcher and are doing a PhD in Education at BU, and Will, you were a senior software engineer at Spotify. How have those two backgrounds combined to help you build QuestionWell so efficiently?
Maya: I think they really complement each other. As a teacher, I can look at the workflows and see what would be most useful and what will get in the way of a teacher’s workflow. Research tends to focus on validity, reliability, and impact, making sure a model is sound and consistently doing good rather than causing harm, and how you even measure that. My PhD is essentially about all of this: how do we get AI to create things that are high quality, and how can it empower teachers to use their own judgment to adjust curriculum to fit their students?
Will: Having worked on ML systems before was definitely helpful. That was something I hadn’t really done before Spotify, where a lot of my job was supporting model serving. Now I’ve had the chance to go deeper into that and do more research-oriented work. Even earlier in my career, I’d worked at some pretty small startups, just a handful of engineers, sometimes a dozen or so, and that was great preparation too. Having seen how successful startups work from the inside, I have a sense of what to look for, even though I’ve got some blind spots.
What’s been your biggest competitive advantage as you have grown?
Will: I think our technical execution has been great, if I can pat us on the back a bit, but Maya’s been a huge part of that too. Having an eye on the research and staying focused on R&D has been our strength so far.
Maya: We were early to this because we really learn by doing. We built a prototype quickly, then looked at feedback from teachers, and it spread by word of mouth. We haven’t done any marketing. It spread because it’s useful. Right now, when a teacher tries an AI tool, they usually have to eyeball the quality themselves. With QuestionWell, our lenses allow the outputs from the models to be compared to specific rubrics so that the quality of each component can be evaluated systematically, and acted on at the moment of teacher decision-making. That’s really tied to our focus on R&D. We’re also a small team, so we talk directly to our customers all the time and adjust the product based on what they tell us.
Maya: R&D for us is both getting input from customers, whether that’s conversations, surveys, or aggregated usage data, and also the research side: taking what we know about good teaching and learning and building it directly into the AI. Everything we’ve talked about, moving beyond just asking ChatGPT and passing along the answer, that’s the R&D part.
What kind of change do you see QuestionWell being a vehicle for?
Will: A lot of professions, including mine, have benefited from technology getting better over time. Teaching hasn’t seen that same benefit, even though teachers are asked to do something hard and with increasing complexity it keeps getting harder. It’s a shame the tools available to them have been so bad, and that’s something we’re working to change.
Maya: The hype around AI has been that it saves teachers time. In practice, teachers just end up spending that saved time going back and forth with AI refining AI-generated materials instead of writing from scratch. So the change we’re after isn’t just “faster,” it’s actually giving that time back: making the job easier, more enjoyable, and more effective at the same time. We’re still early, so most of that opportunity for teachers and students hasn’t been realized yet. Closing that gap is the real goal.
Where do you see QuestionWell heading next?
Will: We want to keep improving our evaluations and our outputs, and keep pushing the state of the art through our own research. It’ll be interesting to watch how things develop. There’s a lot of interest right now in post-trained, open-source models for education, driven largely by concerns around data sovereignty.
Maya: Beyond that, we think about it in terms of three North Stars: quality, agency, and human connection. Quality means the AI produces educational content that’s visibly and consistently pedagogically sound, not AI slop. Agency means giving teachers, students, and administrators more control, and more fine-grained control, not less. Human connection is talked about less, but if every student and teacher is just alone on a computer learning and planning all day, that’s not the outcome we want, so we look for ways AI can actually bring people together.. There’s also a more technical answer to where we’re headed. People often ask, “if a better model comes out, does that make your tool less useful?” For us, it’s the opposite, we’re built on top of these models in a flexible, adaptable way, so as they improve, our product improves right along with them.
Learn more at QuestionWell’s website here.