The stability layer under an AI-heavy real estate product

When Jeremy Bunting joined SERHANT. as VP of Engineering in February 2024, the company's AI product, S.MPLE, was already in the hands of 200 real estate agents. Designed to automate repetitive daily work like market analysis and contact management, S.MPLE was built as a Next.js progressive web app deployed on Vercel — a foundation that let the team expand to a React Native iOS app without touching the backend.

That head start mattered because Bunting faced a challenge familiar to most engineering leaders in the AI space: the ecosystem shifts faster than teams can implement infrastructure updates. The team needed velocity and scale, but also the flexibility to swap models, add capabilities, and adapt as the landscape evolves. The traditional approach would force a choice between speed and flexibility. Bunting wanted both.

An abstraction layer that resists lock-in

As S.MPLE moved from single-model experiments to a production product, the team evaluated Vercel's AI SDK. Bunting's initial concern was lock-in: how much would this tie them to Vercel? One engineer reframed the question — the AI SDK was not infrastructure lock-in but infrastructure independence, an abstraction that hides the complexity of working with multiple model providers.

That distinction mattered. Building tightly around a single frontier model would mean a rewrite with every model change and no clean fallback path when cost or reliability shifted. With the AI SDK, iteration came down to configuration changes rather than feature overhauls. "We are building agentic tools," said Bunting. "Having that consistent abstraction layer across the app for our developers really reduces the cognitive load of understanding how to work with each of the models and services."

AI Gateway added a second layer of leverage: unified visibility into usage across the company's apps and prototypes, even when teams used their own API keys. That consolidated view accelerated debugging and optimization while providing a clear feedback loop on cost.

Routing each task to the right model

Because the team wasn't spending engineering time on one-off AI integrations or infrastructure maintenance, they could focus on testing models against real product tasks and picking the right model for each job:

  • Claude Sonnet for complex, accuracy-critical analysis like comparative market analysis, where structured-data reasoning is key
  • Claude Haiku for lightweight intent detection and field-filling where speed matters
  • OpenAI models for conversational voice and general chat behavior
  • Gemini for image generation, browser automation, and computer-use workflows where reliability and speed come first

The team is also experimenting with using models as guardrails to validate or critique outputs, and with caching strategies to manage token spend as usage climbs.

Scale without replatforming

The soundness of those infrastructure choices showed when S.MPLE launched publicly. "We moved from being an internal pilot program to more than 900 users without a lot of worry on infrastructure or scale," Bunting said. The API layer required no changes, and Fluid compute absorbed the increased workloads automatically.

That scalability matters at SERHANT.'s content pace. "SERHANT. generates about 35% more content than the top five brokerages combined," Bunting noted — property videos, listing descriptions, marketing materials, and now AI-generated assets.

Greg Parsons, Technical Director on the S.MPLE team, said AI Gateway provided visibility that didn't exist before. "We can gain insight into all of the disparate applications we are building across the business, from early prototypes to fully launched products, by bringing everything into a single view," Parsons explained, citing unified visibility into token usage, API calls, and overall system output within one UI.

From linear tasks to conversational agents

S.MPLE started with linear workflows: an agent triggered a single action, it ran end-to-end, and returned a result. Real-world work is more complicated. Users want to execute multiple tasks in one run — producing every asset needed to market a property listing, for example. The team is now building toward conversational experiences where users can run agents, steer and correct them mid-flight, and combine multiple workflows into a single request.

The shift is from one-off automations to a coordinated network of specialized agents that can evolve alongside the AI ecosystem.

Designing for a moving target

Greg Chan, SERHANT.'s CTO, sees flexibility as the core requirement. "In AI, things are evolving fast. What it looks like now is different than even three-to-six months ago. And it'll be different months from now," Chan said. The goal is to keep building within the ecosystem rather than rewriting the stack each time the market shifts. "The last thing we want is to rebuild our stack every time a new model drops," he said.