Inspiration
Learning is not one-size-fits-all.
Two students can face the same concept at the same difficulty level and have completely different experiences understanding it. Most educational software adapts the content — another explanation, another question, a different difficulty — but rarely adapts the actual experience of learning.
That distinction became real when I worked with Roshan, a neurodivergent high-school student studying advanced physics. He was especially interested in improving how he learned difficult topics such as nuclear physics.
My first prototype already had a lot of functionality: predictions, interactive trials, graphs, equations, replay, counterfactual experiments, and AI-driven adaptations. Technically, it worked.
But Roshan identified the problem immediately.
He liked the learning content and interactivity, but told me that the interface felt scattered and did not naturally lead him toward constructing his own answer. More importantly, he did not want the science simplified. He wanted a clearer path through the same rigorous material.
His feedback became the design specification for the final version of UnseenLab.
The goal became:
Structure without oversimplifying. Choice without chaos. Interaction without distraction.
That is where UnseenLab came from.
What it does
UnseenLab turns a learner's STEM question into an interactive learning environment instead of simply returning an answer.
A learner can ask something like:
Why do planets remain in orbit?
or explore a conceptual topic such as cause and effect.
UnseenLab then builds an experience around a structured learning loop:
Predict → Interact → Observe → Explain → Complete
The learner first commits to a prediction. Only then do the relevant interactions become available. They manipulate the model, observe what changes, and explain the result themselves.
Instead of treating AI as an answer generator, UnseenLab uses it to help construct the learning environment.
Depending on the concept, that environment can include interactive 2D or 3D models, diagrams, graphs, equations, text sequences, semantic relationships, simulations, and alternative representations of the same underlying idea.
This also means learners are not limited to a small library of lessons that someone manually authored beforehand. UnseenLab can generate bounded conceptual experiences for additional safe STEM questions while using deterministic engines for supported quantitative simulations.
The central idea is simple:
Same science. Your way in.
How we built it
UnseenLab is built as a layered system rather than a thin LLM wrapper.
A learner's question first passes through a generation and validation pipeline that produces a bounded structured demonstration specification. That specification defines the entities, relationships, controls, representations, lesson prompts, and trust level of the experience.
For supported quantitative concepts, such as orbital motion and Newtonian mechanics, the numerical behavior is owned by deterministic simulation engines. The language model does not get to invent the physics.
For broader conceptual topics, AI can compose qualitative relationships and representations, but those experiences are explicitly treated as conceptual demonstrations rather than verified quantitative simulations.
A canonical concept model connects the different representations. Where applicable, the Diagram, 3D environment, semantic relationships, and guided lesson reference the same underlying entities rather than independently inventing their own versions of the concept.
The 3D system includes deterministic layout, camera framing, semantic object metadata, geometry validation, trajectory rendering, labels, pointer/keyboard interaction, and safeguards against misleading visual relationships.
The learning interface then sits above this model as a state machine:
prediction → interaction → observation → explanation → completion
Progression is tied to real interaction state. Clicking empty space does not count as completing an experiment. Manipulating the actual model or control does.
We also built graceful fallback behavior so the core experience remains usable when model generation is unavailable, rather than presenting a broken AI screen.
Challenges we ran into
The hardest challenge was discovering that more features can make an accessibility-focused product less accessible.
The first prototype had many of the capabilities I wanted, but Roshan's session showed that exposing prediction, graphs, replay, counterfactuals, adaptations, and controls throughout one workspace created cognitive overhead.
That led to one of the largest changes in the entire project: we stopped treating every feature as something that deserved simultaneous screen space.
Roshan specifically suggested making the process sequential — completing one part before moving to the next — while still making the system flexible enough for different ways of learning.
That became the five-stage lesson rail used today.
We also faced a very different technical challenge: making generated interactive environments trustworthy.
A generated lesson cannot tell a student to manipulate a control that does not exist. A 3D causal graph cannot disagree with its 2D diagram. An orbital visualization cannot fake a prettier trajectory than the deterministic engine actually produced.
We built validation contracts around these cases, including control-reference validation, semantic parity between representations, interaction gating, deterministic physics ownership, geometry checks, accessibility paths, and browser-level regression testing.
The final challenge was learning when not to use AI. For verified numerical behavior, deterministic code is more trustworthy than asking a language model to simulate science.
Accomplishments that we're proud of
The accomplishment I care about most is not a test count or a model benchmark.
The final UnseenLab is visibly different because of a real learner.
Roshan's original session identified that the content itself was useful, but the experience lacked structure. He specifically asked for sequential learning, richer visual models, more interaction, and flexibility without oversimplifying the material.
Those requests directly became product changes:
scattered functionality became Predict → Interact → Observe → Explain → Complete basic visualizations evolved into richer interactive 2D/3D representations representation choice became part of the learning experience rather than simultaneous UI clutter AI generation expanded the system beyond a fixed catalog of manually authored lessons interactions became tied to actual model state rather than decorative clicks
In his initial design session, Roshan's self-reported confidence moved from 2/5 before the experience to 3/5 afterward. We intentionally treat this only as a single design case study, not evidence of educational efficacy or a statistically validated learning result.
On the engineering side, we are proud that UnseenLab separates AI composition from scientific authority. AI can help decide how an experience is constructed, while deterministic engines remain responsible for verified quantitative behavior.
That boundary became one of the most important principles of the project:
AI can compose the learning experience without being allowed to invent the science.
What we learned
The biggest thing Roshan taught me was that accessibility does not necessarily mean making difficult material easier.
In fact, he explicitly wanted the opposite.
He wanted the constructive, challenging content to remain. What needed to change was the path through it.
That completely changed how I thought about personalization.
Personalized learning does not have to mean generating an easier explanation for every student. It can mean changing the representation, sequencing, interaction model, pacing, and way a learner explores the exact same rigorous idea.
I also learned that user involvement has to happen early enough to hurt.
Roshan's feedback forced me to throw away assumptions I had already implemented. That was much more valuable than asking someone to approve a finished interface.
Finally, I learned that responsible educational AI often means deliberately limiting the AI. Some parts of a learning experience benefit enormously from generation. Other parts — especially scientific quantities and deterministic behavior — should remain constrained and testable.
What's next for UnseenLab
The next step is expanding UnseenLab's library of deterministic engines while keeping the generative conceptual system available for topics that do not yet have verified simulations.
I also want to continue working directly with more neurodivergent learners across different ages, subjects, and learning preferences. Roshan shaped one major redesign, but one learner cannot represent every neurodivergent experience.
Longer-term, the goal is for UnseenLab to learn which representations genuinely help an individual learner — not based on a diagnosis label, but from the learner's own choices and interactions over time.
A student who understands relationships through diagrams should not be forced into 3D. A student who learns through manipulation should not be handed another wall of text. And a student asking about a topic outside a prebuilt catalog should not simply be told that no lesson exists.
The broader vision is not easier science.
It is more ways into the same science.
Log in or sign up for Devpost to join the conversation.