Hackathon · Tokens LDN Multi-Agent Hackathon · London 2026
Basket
Catch a recipe backlash in week two, not in the quarterly review.
A five-agent pipeline that tells a brand, with sources, when a change to a product's recipe sets off a spike in complaints online.
Best Use of Tavily, 1st place - Tokens LDN Multi-Agent Hackathon 2026
First place in the prize for the best use of Tavily's search API.
- Event
- Tokens LDN Multi-Agent Hackathon, London
- Result
- Best Use of Tavily, 1st place
- Stack
- Python, FastAPI, Tavily, Prometheux, ClickHouse, Next.js
- Team
- Nicole
Why it matters
When a brand quietly changes a recipe, the people who notice say so quickly, in news stories, forums and reviews. Inside the company the usual signal is sales, and a dip there can take a quarter to show up and be traced back to the change. Basket watches the public side instead. Give it a product name and it finds when the recipe changed, collects the complaints, counts them week by week and publishes a sourced alert when they spike.
How it works
Orchestrator
Takes a product name, runs the agents
FastAPI · GET /run
Tavily
Live search of news and the web
advanced depth · dated results
Date finder
Works out when the recipe changed
news from the last 270 days
- three news searches
- earliest date kept
Retriever
Collects complaints, one per link
news and web · last 180 days
- news
- the wider web
Classifier
Sorts each mention by rule
Prometheux · Vadalog rules
- product matched
- complaint phrase found
- dated after the change
taste complaint
Spike detector
Counts complaints week by week
ClickHouse · one count per link
- weekly counts
- peak against the weeks before
Publisher
A sourced alert, only on a real spike
from 1.2× the weekly baseline
Published alert
Decisions we made
Prometheux rules decide what counts
Each mention runs through rules on Prometheux. A list of aliases ties every way of naming the product to one variant, complaint phrases mark it as taste, texture, packaging or price, and its date is compared with the recipe change. The dashboard shows every rule that fired beside the quote, with a link to the source, so anyone can see why a mention was counted.
Trade-off. The rules only know the phrases and product names we gave them, so a complaint worded another way is missed, and each new product needs its aliases written by hand.
A live backlash, not a famous old one
We first planned to replay a recipe change from 2016, but Reddit blocked both ways we had of reading old threads, so we could not date the complaints. Tavily's news results come with dates, so we switched to a live case, the backlash over Reese's Peanut Butter Cups from February 2026, and checked that the spike showed up in dated results before building on it.
Trade-off. The weekly counts rest mostly on news stories, the results with reliable dates, so a backlash has to be big enough to be reported before it shows clearly.
Gallery
The full engineering is on GitHub.