I have spent over a decade turning science into learning, first as a classroom teacher and science program director, now as an academic specialist producing curriculum-aligned lessons at scale. These four pieces show the problem each one solved and how I built it.
A review-gated workflow in which an AI assistant redraws physics figures as editable SVG in one locked style, and every redraw is accepted or rejected against a fixed set of checks.
Science lessons need diagrams that show exactly what the text describes: a free-body diagram with the right forces, a field with arrows pointing the right way. Searching a library of accumulated images usually ends in a near fit. Image generation alone cannot guarantee arrow direction, labels or proportions. An AI-drawn diagram can also look polished and teach the wrong thing, so the real problem was a standard for deciding what counts as good enough to use.
Senior high school learners working through electricity and magnetism, who rely on each arrow and label being right. A creative team reviews the AI-generated figures before they go into lessons.
Standards owner and reviewer. I wrote the style specification, set up the confirm-the-specification step, defined the acceptance checks, and made the accept or reject call on every redraw.
Editable vectors. The assistant builds each figure as SVG, with exact geometry and a file that stays correctable.
Specification first. The assistant writes a short specification of what it will draw, and I confirm or correct it before any drawing starts.
Locked style. Color is assigned by meaning, so a field, a vector and a charge sign are always the same color. Diagram markers keep an outline ring, and surfaces such as planes drop the stroke and use partial fill opacity. Arrows get thick, deliberate strokes because the information lives in the line.
Validation before scale. The style was tested on one example of each diagram type before any batch run.
Labels. Every element is named in the figure or its caption, and labels stand alone without pointing to other items.
Review. Every redraw is judged against the checks below, and every rejection names its reason.
Handoff and traceability. AI-generated images carry a "for checking" comment, and a script gathers them into per-lesson folders for the creative team. Each figure's ID goes into its alt text, so it traces from the lesson to its source file.
The assistant writes the specification and builds the SVG. I define the rules it builds to and the checks it is judged against. I moved part of the feedback loop onto the assistant so that I review a short description before a finished drawing.
Early redraws showed a recurring failure, contrast: light objects on light backgrounds read too faint, and a dark coil on a dark core with dark arrows disappeared. That failure became a named check and a style rule. Keeping the vector files means a correction never means starting over.
I am actively exploring new instructional design and e-learning development roles. If your team is looking to expand or needs someone to lead complex learning initiatives, reach out at ivanhubert.juan@gmail.com
© 2026 Ivan Hubert C. Juan