feat: transition from Gemini 3.1-flash-lite to a local fine-tuned VLM

[exharmonic]

Currently, GradeOps relies on the Gemini 3.1-flash-lite API for vision-based grading tasks. While effective, this creates a dependency on external network calls, introduces latency, and raises potential data privacy concerns regarding student submission uploads.

Proposed Solution

Implement a local, dedicated Vision-Language Model (VLM) fine-tuned for grading tasks. This will allow offline processing and reduced latency/ request overhead.

Technical Requirements

• Research and select a lightweight VLM suitable for deployment (e.g., LLaVA, Qwen-VL, or similar).
• Develop a custom fine-tuning dataset based on previous grading examples to align the model with GradeOps specific rubric requirements.
• Implement a local inference server/module to handle vision tasks via the existing pipeline.
• Update backend/services/graph.py to support local inference endpoints.
• Ensure the hardware resource consumption (VRAM/RAM) stays within reasonable limits for the target deployment environment. (VERY IMPORTANT!!)