AI Content Studio.

Shikho's academic team makes its own curriculum-true videos and sims. Correctness is enforced, not hoped for.

SEE HOW IT WAS DIRECTED →ALL SYSTEMS

INTERNAL SYSTEM · REAL INTERFACE, PROTECTED DATA · WALKTHROUGH AVAILABLE ON REQUEST

ai content studio · internal · real UI●

AI Content Studio - a generated Solar System lesson playing mid-frame, integrity score 100/100

01 · THE PROBLEMWHAT WAS BROKEN

One classroom animation needed a researcher, storyboard writer, animator, cross-checker, edit loops - days and real money, per topic.

02 · THE SYSTEMWHAT SHIPPED · WHO USES IT

The studio's own How-this-works page - the seven-step video pipeline in plain English

The creation wizard - type, class, topic, prompt

A passing render - the rotating Earth lesson frame

REAL STUDIO SCREENS · THE PIPELINE · THE FORM · WHAT COMES BACK

Type in class, chapter, topic - and get back a finished video or simulation.

A four-role pipeline became a form.

637real generations

119approved for classrooms

$3.71average per generation

SOURCE: Usage dashboard · July 2026

03 · UNDER THE HOODWHAT ACTUALLY HAPPENS

Pick a moment. Watch the calls the system makes that no one sees.

CODETHE MODELA CHECKTHE CLEVER BIT00/08

THE THINKING BEHIND EACH STEP

01Why start from one line? Because making a lesson used to need a researcher, a writer, an animator and a checker - days per topic. The teacher now says what to teach.

02A model has no idea what Bangladesh's curriculum says. So each chapter's pack sets the laws first - what this topic must look like, what it must never, which misconceptions to avoid.

03Fixing a mistake in pixels is the most expensive place to catch it. So the critic hits the script first, hardening it against known habits: invented objects, drifting colours, fake text.

04A render costs four dollars, so three cheap checks run first: predict the final frame, confirm the subject, count the objects. A fraction-of-a-cent vision call beats a wasted four-dollar render.

05Two keyframes lock both ends so the model cannot drift off either one. Then eight seconds loop into a lesson instead of rendering longer - because a longer clip is more dollars.

06A video can look beautiful and be scientifically wrong. So it grades its homework - five frames, five weighted dimensions - turning correct from a hope into a number before anyone looks.THE CLEVER BIT

07The score shows on the video, out of 100. A pass isn't a green light taken on faith - it is a visible number the teacher can read and check for themselves.

08A wrong render refuses itself, no human needed. A physics violation - reversed motion, a wrong-coloured liquid caught in a chemistry demo - caps the science score so hard it cannot pass.

EVERY STEP MINED FROM THE REAL BUILD · NOTHING INVENTED

04 · WHY NOT ONE PROMPTWHAT A ONE-SHOT WOULD HAVE GOT WRONG

"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 reverses motion, invents daughter cells, recolours limewater - 14 axioms exist because a wrong render was watched and paid for.
The knowledge had to be built, not assumed. 55 curriculum packs - grounded in NCTB textbooks, agent-researched and cited - encode what Class 7 in Bangladesh is actually taught.
Every render costs real money. Three pre-render gates, one retry each, renders looped instead of lengthened, a hard kill switch on experiment spend.
The misses don't get wasted. Teacher complaints are minted into new eval cases, and a weekly human audit keeps the AI judge honest.

05 · THE DIRECTIONWHERE THE HUMAN WAS

THE FORK · WHAT THE AI PITCHED VS WHAT I SENT

Days in, the pipeline had the model write each simulation from imagination - every render broke differently. His verdict, 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."

THE MODEL ON ITS ARCHITECTURE · AFTER THE REDIRECT

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**.

WHAT I SENT · THE REDIRECT

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, not a prompt: PhET's eight design principles hard-coded, the model reduced to four typed slots it cannot break.

SESSION TRANSCRIPT · 2026-05-16 · BOTH SIDES QUOTED EXACTLY · [...] MARKS OMITTED SPEECH

Don't patch a wrong approach into submission. Go study how the best in the world standardized it - then make that the system.

49 append-only decision records14-axiom constitutionweekly Friday auditsmisses become eval casesHOW IT WAS RUN, NOT HOW IT WORKS

NEXT SYSTEM Presentation Studio →

▸