In the high-stakes world of enterprise AI engineering, the default assumption is that reliability costs money. Scaling accuracy usually demands larger models, endless retry loops, expensive visual processing, and human-in-the-loop fallback teams.
At Brainbase Labs, Engineering Lead Milind Shah is quietly proving the opposite.
Ahead of his Wednesday afternoon session at AI Engineer NYC (“Hill Climbing Back Office Agents”), Shah sat down with The Daily Context to unpack a case study with Domino’s Pizza that flips the standard AI economics curve on its head. Facing a costly, manual process to track competitor pricing and promotional discounts across local fast-food chains, Domino’s deployed autonomous browser agents managed by Brainbase.
Over nine iterative versions, the system’s task accuracy soared from 41% to 91%. Simultaneously, operational costs plummeted from thousands of dollars per month for an external consulting team down to just $20 to $30 per week—with inference costs dropping to $1.16 per source.
The story behind that 50-point accuracy jump offers a masterclass in pragmatic AI engineering for the 1,000+ technical leaders gathered this week at the Sheraton Times Square.
Phase 1: Defining “Correct” Before Writing Code
Before Brainbase touched a single prompt or model parameter, Shah’s team had to solve a foundational problem: defining what success actually looked like.
Domino’s had previously relied on human analysts to manually inspect competitor websites, collect local coupon codes and promo pricing, and construct structured spreadsheets for executive decision-making. To replicate this with software, Shah sat down directly with non-technical business teams at Domino’s to decode their reference spreadsheets.
“They already knew what a good spreadsheet looked like,” Shah explained. “We used that to create an internal benchmark dataset of questions and expected outputs. Without an internal benchmark, you can’t prove accuracy numbers or tie improvements to cost reductions.”
This benchmark became the foundational ground truth against which every prompt tweak, model swap, and agent instruction was scored.
Phase 2: The Coordinate Hack & Model Optimization
Once the benchmark was established, Brainbase systematically tackled the agent’s execution bottlenecks. Surprisingly, the single largest jump in accuracy didn’t come from a bigger LLM—it came from stripping away visual inference.
In early iterations, the browser agent relied on visual processing to locate buttons and navigation links on competitor websites. This approach was slow, expensive, and prone to misclicks. Shah realized that fast-food ordering interfaces have static UI layouts. Instead of making the model parse image tokens to find a “View Deals” button on every page load, Brainbase hardcoded fixed X/Y screen coordinates for recurring UI elements.
“Instead of having the agent figure out where the button is by reading the image, we told it the exact X/Y coordinates,” Shah said. “That eliminated expensive image-reading inference and single-handedly produced our biggest accuracy gain.”
Concurrently, Shah experimented with model selection. Rather than defaulting to Anthropic’s flagship Claude Opus 4.6, the team discovered that tightly specified, highly structured prompts paired with Sonnet 5 delivered superior accuracy at a fraction of the token cost.
Phase 3: The Unsolved Frontier of Self-Improvement
Brainbase’s platform now powers more than 10,000 internal agents across enterprise back-office workflows—ranging from invoice exception clearing to inventory transfers. While Brainbase’s automated “Hillclimbing” engine can automatically detect and remediate failure modes like wrong tool calls or hallucinations, Shah remains candid about the limits of current automation.
Complex browser navigation traces still require human engineering eyes. “Automated improvement works for tool calls and hallucinations,” Shah noted. “But browser navigation is still too intricate for auto-improvement agents today. Engineers still have to manually review agent traces to find where browser actions break down.”
For teams working with tight evaluation budgets, Shah offers straightforward advice: don’t overcomplicate your initial evals. “Deterministic eval criteria go a long way at low inference cost—checking for hallucinations, tool-use correctness, and adherence to knowledge bases.”
The Engineer Behind the Agents
Shah’s path to leading engineering at Brainbase spans software engineering roles at DoorDash and Adobe, scaling the code-data business at Merkle from zero to eight figures, and co-founding the evals platform Pipelines.
Despite building autonomous systems that handle complex enterprise workflows, Shah maintains a grounded, reflective view of the AI revolution.
“I’m optimistic overall, but I think about skill atrophy a lot,” Shah admitted. “If I’m using AI to write code today, and tomorrow I’m using it to write everything else, you risk losing those core skills. A good AI engineer still needs strong engineering fundamentals first.”
Milind Shah speaks on “Hill Climbing Back Office Agents” on Wednesday, October 14 at 2:50 PM ET at AI Engineer NYC.