Building Closed-Loop Evals for a Multimodal Agent at Uber Scale

Talk transcript

1. 0:00

   [upbeat music] My name is Jai, uh, and I'm here with Soumya.

2. 0:19

   We are part of the computer vision team at, at Uber, and we're gonna talk to you about a real world production use- Mic's not on. Oh, mic's not on?

3. 0:28

   Okay. Try again. Okay. Don't worry, I'll, I'll manage. Can you hear me now? Um, okay. So we're gonna talk to you today about a real world production use case.

4. 0:41

   Um, and specifically we're gonna dive into how we design the evals and the eval loops. Um, so let's see this.

5. 0:52

   Okay, it's not working. All right, cool. So just before we get into the agent design, um, we're gonna talk about, a little bit about the use case. So our delivery marketplace, Uber Eats, we do about ninety billion, uh, run rate per year at the moment.

6. 1:08

   Um, we-we're adding millions of items to the marketplace, uh, each and every year. Um, sorry, every, every month. We're growing at twenty percent, uh, year on year, and, and we operate in ten thousand cities globally.

7. 1:21

   So not many people actually know this, but our delivery marketplace is just as big as the mobility side on Uber today.

8. 1:30

   Visual content actually plays a really important role for the user experience. So a photo is quite often the first signal that a customer gets, um, that gives them the, that initial impression about a merchant.

9. 1:44

   Um, so a, a good photo can make the difference between someone scrolling through the feed and actually clicking on an item and adding to the cart. And more and more we're seeing different modalities, uh, on Uber Eats, uh, especially video content.

10. 1:59

    But this is a problem. So our smaller independent merchants simply just don't have the level of quality for their photos that reflect what the eater is actually gonna get.

11. 2:11

    And when we speak to our merchants, there are three themes that kind of emerge: lack of time, lack of know-how, and costs. 'Cause these professional, um, photo shoots actually cost a lot of money, and this can be especially problematic if the merchant is updating their menu over time.

12. 2:32

    So this problem is actually pretty challenging to solve for at scale, right? Because our consumers, they want authentic, real-looking photos. Um, but a meaningful fraction of our consumers actually distrust anything that is AI-generated.

13. 2:48

    So if you open up the Uber Eats app, the last thing that you want is to be scrolling through, uh, you know, food photography that looks like AI slop.

14. 2:56

    So we're threading the needle here. We need to be able to stay faithful to the original image, preserve the brand of the merchant, and avoid everything looking the same.

15. 3:06

    If we have the same prompt for every photo that we're editing, the diversity of the marketplace is gonna collapse.

16. 3:16

    We also, because we operate globally, we also have this long tail distribution of different quality that we see a-across the marketplace. Um, so we've got some examples here. You might see food photography that, you know, has poor sharpness, poor composition, not centered, uh, or, or poor colors as well.

17. 3:35

    We also have a wide range of spectrum of user-generated content on the platform as well.

18. 3:42

    So what are our goals when we're designing these agents? When you think through these goals, you might actually be thinking through, you know, your own agents that you're building yourself.

19. 3:50

    But for us it's about, one, preserving authenticity and trust. Two, improving the quality when we need to, so we wanna be able to improve qu-quality selectively. We wanna optimize globally for, for the entire marketplace.

20. 4:03

    We don't wanna canni- cannibalize certain merchants. We wanna ship safely, and this is gonna be an important theme throughout the talk. We want to learn continuously, and we wanna operate at scale in a cost-efficient manner.

21. 4:18

    So agents are actually really well-suited to solve this problem. So if you imagine a spectrum, on the one side you've got something that's more deterministic, it's more rules-based, um, and, uh, you, you, you have more control over it, but it's fairly, it's a brittle system.

22. 4:36

    It's not actually gonna be able to scale for the entire marketplace. Imagine the other side. You provide an agent with obviously a lot of creativity, it has a lot of agency, um, and that's actually what we wanna lean into.

23. 4:49

    But we can't leave that unconstrained, right? Because we have certain safety and certain guardrails in place that we need to adhere to. So we wanna find a balancing act, uh, and that's kind of set the principle for the way that we think and design around agents and evals.

24. 5:05

    So now we're gonna actually like dive a little bit deeper into a simplified but representative example of what we have in production, and we're gonna go through each stage and how we eval it, and then talk through some continuous learning loops as well.

25. 5:21

    So first up we have what we call an image understanding and routing agent. So this is where multimodality is, is pretty important. We actually ask the LLM to describe what it sees in the photo.

26. 5:33

    Um, and then we, we create a structured output from that and we send it to a router. The router will then determine do we enhance it or do we skip it?

27. 5:42

    If we skip it, we will keep the original. If we enhance it, we send it to our next agent, which is an image editing agent. And this can actually run in a loop, so it gets feedback from a QA agent.

28. 5:56

    Um, it can edit, uh, uh, uh, in, in this loop and self-correct and fix things, um, as it goes.

29. 6:04

    If it goes through a number of loops and it still fails, we, we don't publish it. Then we actually send it to a final post-processing and QA step. If that's all good, we'll publish it to the menu.

30. 6:17

    And the last thing that's really critical is we log everything.

31. 6:23

    Just a quick note about logging. Don't know if you can actually read the JSON here, but you might notice that all of the agents in this end-to-end orchestration is within one-- it's, it's basically a flat structure in this JSON.

32. 6:37

    Um, and so this is actually incredibly useful for the entire team because anyone, be it non-technical or technical, um, focused on engineering product, can actually dive in, um, and look at specific cases to diagnose and also roll up things to look in aggregate.

33. 6:54

    Um, and it's important to note here that, you know, we think this is important to start with. You wanna start with your logging, 'cause if you don't start with it, you have nothing to optimize for, let alone set up a self-learning loop.

34. 7:07

    And at Uber, we, um, we use Arize.

35. 7:11

    Cool. We're gonna dive, um, a bit deeper into the router.

36. 7:17

    So the router's actually pretty straightforward. If you remember, we, you know, we have this multimodality input. We look at certain text description, metadata, the image itself. We ask it to, to, to, to under-- um, describe what it's seeing.

37. 7:29

    We create structured output from that. With that structured output, we can then grade against a rubric. So we have these pass and fail criteria. The last step is we wanna decide whether or not we should enhance or skip.

38. 7:43

    How do we actually eval this? This is-- You can think of this as a more sort of traditional classifier. So here we, we have a confusion matrix. You know, many of you are probably pretty familiar with this.

39. 7:54

    Um, but we can look at things like the true positive cases, the false ne-negative cases, and so on and so forth. Essentially, what we're doing is we're measuring the precision recall.

40. 8:03

    In practice, your routers might actually be much more sophisticated. So for example, we might wanna route an image to a lower latency, smaller model to be able to save on cost and improve the user experience at the trade-off of quality.

41. 8:17

    And i-if that's the case, instead of having a two-by-two matrix for your confusion matrix, you might actually have an end-by-end matrix, where each grid is actually telling you whether or not you're correctly routing to that specific branch.

42. 8:32

    So I'm gonna now, um, hand over to Soumya, who's gonna dive a little bit deeper into how we handle drift and human alignment.

43. 8:41

    Do I-- Do I take the clicker as well?

44. 8:45

    Okay.

45. 8:48

    So now that we spoke about how we eval the routing, I wanna talk about how do you get the first version of the model out. For our use case, we consider human labels as the golden source of truth, and this is what we want to align our models to.

46. 9:01

    The way we do about this is we go collect a dataset which is representative, so, you know, different cuts, geographies, dish type, image quality type, send it to our human labelers, and give them a very objective guideline to label on.

47. 9:14

    This is to remove any subjective biases or any noise coming in from human labelers. Once we've got that system set up is when we start tuning our model. We take our agent, we go ahead and get output from the agent, compare it to your golden dataset, evaluate if it's good enough to ship.

48. 9:28

    If it meets your guardrail metrics, you go ahead and ship it. If not, then you go tune and keep doing this until you meet your guardrail metrics. For routing, our guardrail metric is recall.

49. 9:38

    We don't want any bad image to slip through our system.

50. 9:44

    Here are some examples of the failures we've seen. Um, on your left, you see a very good image of cheeseburger. Um, on the right, you notice that the routing agent actually failed this.

51. 9:53

    It said the technical is low bal, and it'll go send this image for enhancement. Now, there's two challenges when you send this image for enhancement. Firstly, you pay the compute cost for a zero quality lift from this image.

52. 10:04

    And secondly, uh, there is a risk of degrading this image, given it's already such a high-quality image.

53. 10:13

    And on the other end of the spectrum, you have a recall miss. So on your left, you have an image with six chicken wings, and on your right, if you notice the dish name, it says eight pieces chicken wings.

54. 10:23

    And your routing agent approved this image, that means. So now there's a risk here. If you send up-- send this image for enhancement, and you only see six chicken wings, there's a chance your model is gonna hallucinate these two extra wings just to match the description.

55. 10:36

    And that's also an, uh-- That's a, the co-- um, cut we take at our faithfulness metric that Jai earlier showed us. So the meta point I'm trying to get here is you've trained your offline model, but there will be long cases where your model is going to continue to fail, and the static model will not work in

56. 10:51

    the real system. You need a way such that your prompts, agents, system itself is evolving over time. And that's what we've done, uh, for our system as well. And I'm talking more from the routing perspective, but every component in our system is able to tune itself, uh, for any drift online.

57. 11:08

    So what we do is we sample production data at regular cadence, uh, send this to the human labelers with the same guidelines that we have se-seen before. Once you've got that data, we compare our agent's output with the output we got from the labelers and see if there's a mismatch.

58. 11:22

    If there's a mismatch, we have an umbrella diagnoser agent, which takes in the feedback, localizes where this issue is happening, and sends-- and triggers our autotuning pipeline. Once we tune this agent, we go and benchmark it against our golden dataset that we saw earlier.

59. 11:38

    And if we pass our golden dataset on the metrics that we had designed, we go ahead and ship this model. Uh, if not, then you kinda keep iterating. And this happens on a regular basis on production dataset.

60. 11:49

    Um, the beauty of this is this is completely config-driven and doesn't require human-in-the-loop. Your diagnoser agent can write your config and trigger the autotuning pipeline here. And this is what will keep your models sharp over time.

61. 12:01

    You will have one static model with the offline, but this is what is going to keep your system alive.

62. 12:08

    Um, so Jai is gonna spend more time on the diagnosis side of it. What I want to do is zoom into the auto-tuning bit. And again, we are looking at routing, but this is how we tune every agent in our system.

63. 12:20

    Um, so we start with a target agent, and we've already got these, uh, unseen eval samples from our humans. We go find out the mismatch and matches and call a prompt optimizer agent.

64. 12:30

    Now, this itself is two sub-agents. There's the ref-reflect agent and the op-- synthesize agent. What reflect does is it, it just looks at the mismatches, tries to find... remove any noise, find any systemic issues that might be in your dataset, and reflect on it and send that feedback to the synthesize agent.

65. 12:48

    Now, the syn-synthesize agent takes this feedback. It has your agent config. It goes and updates your agent with the new config based on the feedback it's getting and goes and benchmarks again.

66. 12:57

    If this benchmark is passed, you actually register this new agent in the new agent config store, and next time your production runs, you pick up the new version of the agent.

67. 13:07

    And this is a c-closed loop system, as I mentioned, no human in the loop. We definitely have observability on the guardrails. Quick rollback built in in case of, uh, any issues with the system itself.

68. 13:20

    Moving on to the next step of our orchestration flow. So we spoke about routing, moving on to the enhancement bit of it. It's a three-step process. What we do is the first step, we generate a prompt specific to this image.

69. 13:32

    We take in the this description, we take in the directives we were getting from our routing agent, and we go ahead and generate a prompt for this image, what needs improvement in this image specifically.

70. 13:40

    And we go ahead and enhance this image. Then you've got the QA gate, which is a multidimensional gate, looks at multiple things like plating, faithfulness, colors. And it, if it passes is when you actually go ahead and publish this.

71. 13:52

    If it doesn't pass, you take the feedback back from the QA gate, push it back to your generate prompt along with the initial inputs you sent it, and go ahead and enhance it again.

72. 14:00

    So there's two end results here. You either keep enhancing for K iterations, and you pass your QA gate, and you publish, or you take a coverage hit, and you never enhance this image.

73. 14:11

    Here's an example. On your left, you see a bowl of sweet potato fries. We send it, uh, for the first iteration, and our QA agent rejects it because the portion size is in-incorrect.

74. 14:20

    The plating is very unrealistic. We take that feedback in, go for the second iteration, and we're actually able to pass it the second iteration. So the metric we are measuring here is pass at K.

75. 14:30

    Pass at K is essentially the pass rate at K at iteration. And ideally, with the more the iterations, your, uh, pass rate will increase because you're getting more feedback in.

76. 14:40

    Now I'll pass it on back to Jai to cover the rest of this.

77. 14:47

    Thanks. Thanks, Soumya. Um, so yeah, just before we end here on the, um, on, on the generation evals, we use what's called pairwise comparison, right, for our pass at K.

78. 15:01

    So it's looking at the input image and the, the output image, and it's assessing whether or not it's better. But how do we actually find what's better? So, um, we're not gonna dive into too much of the details here 'cause this is kind of like proprietary stuff.

79. 15:16

    And so we'll just mention at a high level that this is where you sort of, for at least for us at Uber, we have to make sure that we're aligning with product, design, policy, legal.

80. 15:26

    And this is where we're baking in what we define as a better image on the platform into our evals. Um, so examples here, is it faithful? Is it complete?

81. 15:36

    Is it natural? Is it realistic? And there's a bunch of other things as well. The output of this is then, uh, a yes, no, or unsure.

82. 15:45

    So here are some examples of failure modes. So input and output on the right. The input's on the left, output's on the right-hand side. This might be a little bit, uh, difficult to, to see at a first pass, but we actually added shrimp here, and we shouldn't be.

83. 15:59

    So we fail faithfulness. This is where we go the other way. So the input, um, has some sauce at the bottom of the sushi. We actually remove it. So we fail completeness.

84. 16:15

    Here's actually a, a pretty interesting example where the agent actually attempted a more creative edit the first iteration, um, and then the QA said, "Nope, that's not good enough."

85. 16:26

    Uh, and then it actually oversteers the other, other way, and it becomes overly conservative, sort of falls back to this generic ceramic plate, uh, ceramic bowl, sorry. So this is an example of a reward hacking actually.

86. 16:40

    And, and this is a nugatory change, but it's something that we don't think is a meaningful or influential change despite the actual raw pixels of the input and output being pretty different.

87. 16:50

    Here's another example where in the output, the plate is covering the sauce. This is an example where for... So some of the frontier models that we're using for the actual image editing, some of their, um, some of their, um, problems will actually sort of leak up into our applied use case.

88. 17:06

    Um, and so, so object coherence and ph-physics plausibility are the evals that sometimes we'll coordinate with the frontier teams and, and let them know about these problems and work together with them.

89. 17:18

    Here's, uh, an example of why multimodality is, is pretty important. In the input and the output, we, we can't actually see that there are eight pieces here of, of the wontons.

90. 17:28

    So we're not confident actually. We're not sure. And so this is an example where we would actually reject it in production and, and it wouldn't, it wouldn't go through.

91. 17:39

    So the last step after all of that is a post-processing and what we refer to as the publish-ready QA. This is the final gate before we decide we wanna publish something to production.

92. 17:53

    Here we do some policy checks. We also do some more quality checks. Um, and you might be wondering like, "We've already done some QA, like, why are we gonna do another step of QA?"

93. 18:04

    The reason is because we think of this like a Swiss cheese model. So we want to try and optimize for reducing the chance of a failure getting into production.

94. 18:15

    And so there is some redundancy here or there, and that's okay. Um, and so this QA gate is, is a little bit more holistic. It captures more things, but it also will, will try and flag things that we should have caught upstream as well.

95. 18:33

    All right, so we've talked about a, a couple of, uh, feedback loops here. So to summarize, we talked about predominantly this first one here, which is the model loop.

96. 18:44

    And this is accounting for drifts and aligning with human label dataset that we have and we've established offline. But we actually have more feedback loops. So we, we have at Uber, what we have is a, is a great sort of dog, dog fooding culture, um, where we'll test apps before they go live.

97. 19:03

    Um, but we also have when it goes live in production, how do we get that feedback back into our agent to be able to steer it appropriately? So as we're adding more of these feedback loops, we want to be able to generalize the system.

98. 19:17

    So this is where we've actually created, um, a higher level of abstraction on top, which we call the diagnoser. So the diagnoser can take in any input from these different feedback loops that we're, we're capturing.

99. 19:29

    It can reflect on what actual agent within the overall system needs to be optimized, and it can route to that agent to be able to fix that configuration specifically.

100. 19:40

     It could be one agent, it could be multiple agents.

101. 19:45

     So here's an example of internal dog fooding. You might see these in sort of different apps that you've got where you've got the thumbs down and the thumbs up.

102. 19:52

     We also take some free form feedback as well. Uh, and this is actually great 'cause we'll get feedback from merchants directly. We'll get feedback from, you know, design teams, other product teams, uh, at Uber, and we'll incorporate that feedback back into our diagnoser step and tune the system over time.

103. 20:11

     Again, similar sort of workflow h- pattern here. We'll replay the examples that we know are those ones that have been flagged, be it good examples, be it bad examples, uh, and then we'll benchmark the metrics before we push the latest config version.

104. 20:27

     The last step is, is actually getting this into production and, and this is where we're looking for a whole heap of different metrics that we track for, for the marketplace quality and health.

105. 20:38

     Uh, I've just called out one here, which is conversion. So we're looking for improvements in people adding to cart, converting, completing their orders. Um, I think this one's actually an interesting one to call out because now, I mean, at least at Uber, but especially in production, um, settings at scale, you have a wide, um, uh, you, you

106. 20:57

     have a lot of data that you can actually slice and dice. So in this area as opposed to the others, what we can do is sort of slice by geos, by device type, by dish type, et cetera.

107. 21:08

     And we can look at where things are improving, uh, in different segments and actually tune on certain segments as well.

108. 21:17

     Cool. And that's it for our presentation. Appreciate it. [audience applauding] [outro music]