A guided video lesson
Closed-Loop Multimodal Evals: From Routing to Marketplace Feedback
A source-grounded lesson on building a production multimodal image-enhancement system with selective routing, bounded editing, layered QA, offline human alignment, online drift correction, and marketplace feedback.
Chapter pages
Choose where to begin
- 01Scope: a production multimodal eval problem0:19 - 0:56
- 02Why food imagery matters at marketplace scale0:56 - 2:01
- 03The merchant-side quality problem2:01 - 2:34
- 04Threading the needle: authenticity, faithfulness, and diversity2:34 - 3:18
- 05Long-tail input quality and concrete defects3:18 - 3:43
- 06Design goals for the agents3:43 - 4:23
- 07Why agentic flexibility needs guardrails4:23 - 5:06
- 08The representative end-to-end example5:06 - 5:23
- 09Core orchestration: understand, route, edit, and gate5:23 - 6:19
- 10Logging as the foundation for diagnosis6:19 - 7:13
- 11What the router consumes and decides7:13 - 7:45
- 12Evaluating routing with binary and multi-branch metrics7:45 - 8:35
- 13Human-aligned offline tuning and the recall guardrail8:35 - 9:45
- 14Failure mode: sending a good image for enhancement9:45 - 10:14
- 15Failure mode: a recall miss and cross-modal hallucination risk10:14 - 10:45
- 16From static offline models to online drift correction10:45 - 12:10
- 17Prompt optimization, versioning, and safe closed-loop deployment12:10 - 13:21
- 18The bounded enhancement-and-QA loop13:21 - 14:13
- 19Worked example: sweet potato fries and pass at K14:13 - 14:49
- 20Pairwise comparison: deciding what counts as better14:49 - 15:47
- 21Failure modes: faithfulness, completeness, and reward hacking15:47 - 17:20
- 22Multimodal uncertainty should block publication17:20 - 17:42
- 23Publish-ready QA as a redundant final defense17:42 - 18:35
- 24Several feedback loops, not just the model loop18:35 - 19:19
- 25The diagnoser: a higher-level feedback abstraction19:19 - 19:47
- 26Dogfooding, human feedback, and regression replay19:47 - 20:29
- 27Production outcomes and segment-specific tuning20:29 - 21:36