Soumya Gupta and Jai Chopra describe a multimodal pipeline that routes merchant photographs for enhancement or leaves them untouched. Its goals include better quality while preserving ingredients, portions and each merchant's identity, rather than making every photograph look the same.
Human-labelled datasets spanning dishes, regions and image quality calibrate routing precision and recall. Production samples reveal drift, and diagnostic, reflection and synthesis agents propose configuration changes that must pass existing benchmarks before promotion.
Image-specific editing prompts receive limited quality-feedback retries. Pairwise checks assess faithfulness, completeness and realism: invented shrimp, missing sauce and superficial overcorrection remain failures. A separate publication check can catch errors missed upstream.
The presenters combine human reviews, merchant feedback and production outcomes to target weak components. Their approach prioritises observability and bounded, benchmarked updates rather than an unconstrained self-improving loop.
Watch on YouTube
