Skip to main content
MediaHomework gradingMathLecture videoHandwriting OCR

A graded math worksheet, then a lecture video for the mistakes

Handed a photo of a child’s summer math worksheet, Xiaohe read the handwriting, graded all ten visible problems, marked the two wrong answers, and produced a 6-minute voiced lecture that explains exactly how to solve them.

Watch the lecture

6:00 · 1920×1080

From the photo to the lecture

The task

A user handed Xiaohe a phone photo of a child’s summer math worksheet — ten problems, each with a handwritten answer in the blank. The ask: grade it, then turn the two wrong problems into a short lecture video.

Reading the handwriting was the first hurdle. The photo is a large HEIC, the page is photographed sideways, and the answers are pencil-thin. Xiaohe rotated the page upright, then used a vision model together with local OCR to transcribe every printed problem and every handwritten answer.

Xiaohe solved all ten problems and found two errors: Q9’s second value (45 should be 49 — the two cases need both boat speeds) and Q10’s first blank (240 should be 16 — 240 is the total gap at minute 15, not the per-minute rate). It marked the photo, then built a 13-slide lecture and a 6-minute voiced video walking through both problems for an elementary student.

How Xiaohe did it

  • 1. Read the photo

    HEIC → portrait orientation; Apple Vision and qwen-vl transcribe the printed problems and the handwritten answers.

  • 2. Grade

    Solve all ten problems; settle two ambiguous handwritten readings by inspecting tight crops pixel by pixel.

  • 3. Mark the scan

    Green ✓ on the correct answers, red circles plus the correct answers beside the two mistakes.

  • 4. Build the lecture

    lecture-agent writes 13 HTML slides — river and track diagrams, a timeline, step-by-step arithmetic in large type.

  • 5. Narrate

    Doubao TTS reads a 6-minute walk-through written in plain language for elementary students.

  • 6. Assemble

    HTML → PNG frames and SVG → PPTX; PNG frames + page audio → 1920×1080 MP4.

  • 7. QA and deliver

    Rendered the PPTX to catch text-box overflow, then shipped the video, deck, and graded image back to the source folder.

Skills used

  • lecture-agent
  • ppt-agent · HTML → SVG → PPTX
  • Vision OCR (qwen-vl + Apple Vision)
  • Doubao TTS

Deliverables

More works

Real tasks, from the terminal to a finished document.

All works