AI Content Studio.
Shikho's academic team self-serves curriculum-true animated videos and interactive physics simulations. Governed by a 14-axiom academic constitution - correctness is enforced, not hoped for.
INTERNAL SYSTEM · REAL INTERFACE, PROTECTED DATA · WALKTHROUGH AVAILABLE ON REQUEST
One classroom animation used to need a researcher, a storyboard writer, an animator, a cross-checker, and edit loops between them. Different people, days of coordination, real money - per topic.
REAL STUDIO SCREENS · THE PIPELINE · THE FORM · WHAT COMES BACK
The academic team types class, chapter, topic - and gets back a curriculum-true animated video or an interactive simulation. 637 real generations at $3.71 average; 119 approved for classroom use so far.
What changed: a four-role pipeline - researcher, storyboarder, animator, checker - became a form.
SOURCE: Usage dashboard · July 2026
Pick a moment. This is what the system does with it - every step a deliberate call, none of it visible from the outside.
"Make educational videos with AI" is a one-prompt idea. Making them curriculum-true at four dollars a render is not.
- Stochastic models do not respect physics. Veo happily reverses motion, invents extra daughter cells, and recolours limewater. Every axiom in the 14-rule constitution and every named trap exists because a specific wrong render was watched, and paid for.
- The knowledge had to be built, not assumed. 55 curriculum packs grounded in the actual NCTB textbooks - researched by an agent swarm, cited, version-controlled - give the system the one thing no model reliably knows: what Class 7 in Bangladesh is actually taught.
- Cost discipline is a design layer. Three pre-render gates each capped at one retry, renders looped instead of lengthened, a hard kill switch on experiment spend - engineering against a real per-render price.
- The failure loop feeds the system. Teacher complaints are minted into new eval cases, and a weekly human audit keeps the AI judge honest. A one-shot has no way to learn from its own misses.
Days into the simulation feature, the pipeline asked the model to write each simulation as a full HTML app from imagination - and every render invented a new class of bug. His verdict on the output, verbatim: "Lastly, for the lack of a better word, the simulation looks like absolute shit.
PhET Interactive Simulations is like the gold standard, they have so many great simulations - and a lot of it is open to use."
You're right - the fix loop is fighting the wrong battle. The root cause is non-deterministic scaffolding from Gemini; standardizing the scaffold from PhET's actual practice eliminates the class of failures we're patching after the fact. [...] Got the research. Calling it: **Option B (mini-PhET runtime) shipped as the deterministic shell, with Gemini reduced to filling slots**.
I think we are overcomplicating the code generation part of the simulation - that is causing the root problem, which we are then struggling to fix.
A better approach might be to digest - https://github.com/orgs/phetsims/repositories [...] /goal - ingest the simulations, truly deeply study them, find out and create our systematic version of this so that we have a more deterministic system rather than based on whatever the prompts give us.
The playbook is now code: PhET's eight design principles hard-coded into the design pass, a deterministic mini-runtime owning everything structural, and the model reduced to four typed slots it cannot break.
Don't patch a wrong approach into submission. Go study how the best in the world standardized it - then make that the system.