Assignment Solving Engine

The orchestrator is a planner-executor loop. Given a prompt (text, PDF, or image), it produces a plan, executes each step with the right skill, and assembles the output in the format you asked for.

Inputs

• Text prompt: the assignment brief, pasted into Telegram.
• PDF: parsed with pdf-parse, chunked if large.
• Image: routed through Tesseract.js OCR; passed to the LLM as both extracted text and the original image (for multimodal models).

The pipeline

# simplified trace for a typical assignment
1. ingest        → normalize prompt + attachments
2. plan          → LLM produces JSON plan of sub-tasks
3. dispatch      → orchestrator routes each sub-task to a skill
4. solve         → theory / code / math / citation skills run
5. format        → render to PDF | code file | freeze screenshot
6. package       → bundle artifacts into a single submission

Example plan

Scuba asks the model to emit something like this:

{
  "course": "CS2003 — Data Structures",
  "title": "Lab 04: Linked List",
  "deliverables": [
    { "type": "code", "file": "linked_list.py" },
    { "type": "pdf",  "sections": ["Introduction", "Algorithm", "Output"] },
    { "type": "freeze", "source": "linked_list.py" }
  ],
  "submit": {
    "target": "google-classroom",
    "course": "CS2003",
    "assignment": "Lab 04"
  }
}

Skills

• theory: long-form answers, well-cited.
• code: generates + runs code in a sandbox, captures stdout.
• math: LaTeX-first answers, rendered into the PDF.
• cite: adds references when the rubric requires them.

The human-in-the-loop

Before submitting, Scuba sends you a preview of the final PDF and a Confirm / Edit / Regenerate inline keyboard. Nothing gets turned in without your tap.

Next

See Output Formats for how artifacts are rendered, or Google Classroom Integration for the submission step.