Brainbase Labs took a Domino's pricing agent from 41% accuracy to 97% accuracy across 11 versions. MTS Milind Sunil Shah detailed the demo to AI Engineer with no exploit involved. Affected systems were competitor websites, delivery apps and internal pricing spreadsheets. Access required local analyst review, not remote attacker access.
Before that climb, Brainbase Labs entered Y Combinator Winter 2024 as The AI Labor Company, per Round Funded. At version 1 the agent ran 129 minutes per run at $2 per source for 41% accuracy, Brainbase Labs said. At version 11 it ran 71 minutes per run at $1 per source for 97% accuracy, Brainbase Labs said.
Brainbase Labs started with Domino's 20,000 plus stores in 90 plus markets. The CMO office wanted the lowest price for each item on every nearby competitor site and app. Brainbase Labs said analysts spent 2 hours daily building the spreadsheet by hand, and one wrong step left a finished-looking cell that was just wrong.
Brainbase Labs built a six-step loop to fix it: kickoff, anthology, benchmarks, first agent, iterating, and handoff. Shah said version one's job is to be measured, it is not to be good. Brainbase Labs shipped a narrow first agent for one app, one item, one city, then widened to burgers, takeaway and more sites once the baseline existed.
Anthology meant every term got a definition. Brainbase Labs stored targets in a Postgres-like memory, told Brainbase Labs' agent to record the checkout price not the strikethrough price, and to flag gaps not guess. Benchmarks reused old analyst spreadsheets and added stored screenshots when archive.org failed for apps, while evals used code checks, LM as judge for traces and agent as judge for flags.
The daily brief
Get the free daily brief with the day's AI news, new AI startups and funding rounds every morning.
Join 64,000 other subscribers.
Iterating did the work. A missed $21 deal below the linked pizza section forced a rule to scan the whole page, Brainbase Labs said. A Brand D promos page with 26 image-only prices sent the browser agent into 15 minutes of zoom loops, so Brainbase Labs told it to read HTML DOM, titles and alt text first and stop past a zoom threshold, cutting image reads by one third and cost by over 50%, Brainbase Labs said.
Brainbase Labs now hands off the loop, not just the agent. Analysts review only exceptions, each correction becomes a new benchmark case, and the agent runs on a daily schedule or by email trigger, Brainbase Labs said. Brainbase Labs said expansion to additional Domino's markets and use cases is next.
