Sr. Full-Stack Engineer, Data Systems

Datavations is a leading New York-based data and AI software specializing in the $2.3 trillion dollar building materials industry. Datavations gives building materials and home improvement manufacturers real-time, store-level visibility into pricing, assortment, and inventory across major retailers, including Home Depot, Lowe’s, and Menards. Manufacturers use this data to make sharper decisions about pricing, distribution, and how they show up on the shelf.

Product taxonomy and attributes form the foundation of every insight Datavations delivers to our customers. We are looking for an engineer to own this domain end-to-end—driving both the data systems that organize the market and the applications that enable our teams and customers to interact with them. This is a hands-on role with significant architectural latitude for someone who wants to take full ownership of a system that already powers the business.

• The attribute extraction engine — LLM-based extraction at production scale, its configuration model, its quality gates, and its cost profile.
• The transformation layer for this domain — the dbt models that turn extracted values into published attributes: standardization, and the override model that lets human judgment reliably beat the machine. Built on our Click House warehouse alongside the data platform team.
• Orchestration and reliability for these pipelines — the Dagster jobs, sensors, and schedules that run taxonomy and attribute processing, including run monitoring, retries, alerting, and recovery tooling. Reliability is not a separate team here; for this domain, it is this seat.
• The applications — internal and customer-facing. You own the screens people actually use, not only the services behind them: the internal platform our teams run taxonomy and attributes from, and the customer-facing views of this data as they move onto it. React/Next.js front end, backend services and APIs, application architecture, deployment and CI/CD.
• The write path — every change validated, logged, and visible before it ships, with approval where it matters.
• Data quality and observability — automated testing on the data itself, plus statistical detection of what goes wrong quietly: outliers, drift, a retailer that stopped updating, a value that flips between runs.
• Engineering standards for the domain — documentation, tests, runbooks, and review culture as the system and the team around it scale.

• A meaningful part of this platform is AI, and not bolted on the side: an LLM extraction engine already running at scale, a rules engine that produces a confidence score per item, and an in-app assistant that answers questions in natural language and proposes changes for review. We are looking for someone who has built this kind of system properly, not someone who has called a completion endpoint.
• Evaluation before assertion — golden sets, regression suites that run when a prompt changes, and a defensible answer to “did that make it better?” Quality you can measure, not quality you claim.
• Prompts and rules as versioned data — stored, diffable, testable, auditable, with a clear record of what changed and what it did. Not constants in a file that move on deploy.
• Tool-calling and MCP — exposing internal systems to agents safely: scoped, read-first, audited. It is how our teams will query and operate the platform, and how we already work internally.
• Cost and latency as design constraints — model tiering, caching, batching, circuit breakers. Knowing what a run costs before it runs, and why a small model first is usually the right answer.

• A modernization already in motion, not a greenfield and not a rescue: a production system that has grown…