Feature 10: Multi-Model AI Ensemble Engine

[saidai-bhuvanesh]

Description

This issue proposes upgrading the core prediction engine to a Multi-Model AI Ensemble Engine. Currently, Stream A and Stream B outputs are fused via a fixed linear combination.

We will implement dynamic weighting based on input quality metrics: if the eye crop is dark or blurry, its ensemble weight is automatically scaled down and the body/gill weights are scaled up. We will also implement confidence intervals using logit distribution spread, and add fallback logic.

Technical Implementation Details

1. Dynamic Weights: Modify main.py to calculate dynamic stream weights.
2. Confidence Interval: Compute margin of error from probability deviations.
3. Graceful Degradation: Recalibrate weights if a stream fails to load.

[arcgod-design]

Hi @jpdevhub! I'd love to build this as part of SSoC 2026 👋

A Multi-Model AI Ensemble Engine is exactly what FreshScanAi needs to improve reliability — combining predictions from multiple models reduces individual model biases.

My Approach:

• Integrate 2-3 complementary models (e.g., MobileNetV3, EfficientNet, custom CNN) for freshness scoring
• Implement weighted ensemble voting: each model gets a confidence weight based on its historical accuracy
• Aggregate predictions using soft voting (average confidence scores) for the final freshness label
• Display per-model confidence breakdown in the results UI for transparency
• Benchmark accuracy improvement vs. single-model baseline

Could you please assign this to me (@arcgod-design)? Thank you!