Descriptions:
Engineers Soumya Gupta and Jai Chopra from Uber’s computer vision team walk through their production multimodal agent for Uber Eats food photo quality enhancement — a system operating across 10,000 cities and supporting a $90 billion annual run-rate marketplace. The talk focuses not on the image-editing model itself, but on the closed-loop evaluation infrastructure that makes it safe to run at scale.
The pipeline consists of three agent stages — image understanding and routing, enhancement, and quality verification — each with its own guardrail metrics and human-calibrated golden datasets. The speakers detail specific failure patterns: a routing agent incorrectly flagging a high-quality cheeseburger photo for unnecessary enhancement, and a recall miss where a six-piece chicken wing photo slips through despite the menu listing eight pieces, creating downstream hallucination risk. Recall is the primary routing guardrail precisely to prevent these cases.
The most distinctive element is the continuous learning loop: the system samples production data at regular cadences, surfaces mismatches between agent output and human labels, and triggers an automated diagnosis agent that localizes drift, rewrites prompts, retrains, and re-benchmarks against the golden dataset — all without human-in-the-loop intervention. This config-driven self-improvement architecture offers a replicable template for teams building agents that must stay calibrated over time in long-tail, globally diverse environments.
📺 Source: AI Engineer · Published July 24, 2026
🏷️ Format: Workflow Case Study
