Real architecture · Mid
Ailo: LLM recommendation agent (16 s and nondeterminism)
An honest failure (the "cookie" shown to someone who ordered a hamburger, due to a taxonomy error), 16 s or more of latency, and two fixes. The bottleneck was the data, not the model.
Short answer
An honest failure (the "cookie" shown to someone who ordered a hamburger, due to a taxonomy error), 16 s or more of latency, and two fixes. The bottleneck was the data, not the model. Ailo is iFood's AI agent for consumers. According to the post, it has stopped being just a small experiment and is becoming a product.
01 · CONTEXT
What Ailo is
Ailo is iFood's AI agent for consumers. According to the post, it has stopped being just a small experiment and is becoming a product. The person talks to the agent, which presents a carousel of items based on order history and the context of the interaction; then, the journey continues in the iFood app.
According to the post, Ailo is available in the iFood app in five cities and on WhatsApp.
The path began with a simple decision-tree bot on WhatsApp, which allowed orders at specific restaurants. The second experiment focused on basic recommendations of restaurants and dishes.
02 · DATA BEFORE THE MODEL
A recommender lives off its data foundation
The post states that the hardest part of building a recommender is not the model itself, but the quality of the data foundation. It cites iFood's structured base and the LCM, which ensures that user intent is always aligned with the offer options in real time.
The person talks to the agent, in the iFood app or on WhatsApp.
The agent presents a carousel of items based on order history and the context of the interaction.
The LCM keeps the person's intent aligned with the offer options in real time.
After the carousel, the journey continues in the iFood app, in the experience that already exists.
The post also describes a two-way street: Ailo learns from each interaction and improves from user feedback.
03 · THE LIVE FAILURE
The cookie for someone who asked for a hamburger
The development team was on call, watching a crucial demo for international investors. The plan was to show personalized recommendations from a single keyword.
The user asked for a hamburger, and at that hour in Brazil, only one burger place was open at the searched address. According to the post, the agent concluded that the user had not liked the only burger available and went off to run a new search; because of a small taxonomy classification error in the catalog, it presented a cookie.
The person asks the agent for a hamburger.
At the searched address, only one burger place is open.
The agent concludes that the person did not like the only burger available and runs a new search.
Because of a small taxonomy classification error in the catalog, the agent presents a cookie.
The post also attributes the failure to the agent being nondeterministic: unable to find an offer, it looked for a solution at all costs. The team itself acknowledges it did not deliver its best.
04 · LATENCY
Sixteen seconds and two fixes
At first, the agent's latency reached sixteen seconds or more to generate results. The post describes two moves to reduce it: adjusting how each message is processed and showing micro-interactions during complex searches.
At first, generating results took sixteen seconds or more.
Simple requests do not trigger deep searches, which ensures faster responses.
In complex searches, the person sees micro-interactions and notices that Ailo is working in the background.
| Criterion | Simple message | Complex search |
|---|---|---|
| Tratamento | Does not trigger deep search | Shows micro-interactions while searching |
| Described effect | Faster responses | The person notices that Ailo is working in the background |
The post says that, with research and constant monitoring, the time was reduced to acceptable levels, but it does not publish the final value.
05 · CHANNEL AND RESULTS
The channel shapes behavior
In the app, fewer than half of users run more than one search. According to the post, this happens because the comparison is always with the traditional iFood experience. On WhatsApp, the dialogue is more natural and continuous.
The post reports that Ailo users create twice as many carts as in the traditional app experience and that, in the team's surveys, eighty percent of users find Ailo useful in moments of indecision.
A current challenge mentioned is balancing two important moments: exploring new options and the comfort of surfacing what the user already knows and likes.
06 · LESSONS AND LIMITS
What to take into an interview
Among the team's lessons: it is not enough to recommend; you need to understand the moment, the time of day, the mood, and the story. And the person's final autonomy is not replaced by the agent without their permission.
Frequently asked questions
What is Ailo: LLM recommendation agent (16 s and nondeterminism)?
An honest failure (the "cookie" shown to someone who ordered a hamburger, due to a taxonomy error), 16 s or more of latency, and two fixes. The bottleneck was the data, not the model.
What does the Ailo: LLM recommendation agent (16 s and nondeterminism) design cover?
It is organized in these sections: What Ailo is, A recommender lives off its data foundation, The cookie for someone who asked for a hamburger, Sixteen seconds and two fixes, The channel shapes behavior, and What to take into an interview.
How long is Ailo: LLM recommendation agent (16 s and nondeterminism) and what level is it?
It takes about 15 minutes to read and is rated Mid level.
Now design it yourself
Run this case as an interview in the System Design school: you make the decisions, the tutor challenges them and you see how your design scores.