Principal Applied AI Engineer - Entity Agents

ZoomInfo | Product | Core Data

ZoomInfo is where careers accelerate. We move fast, think boldly, and empower you to do the best work of your life. You’ll be surrounded by teammates who care deeply, challenge each other, and celebrate wins. With tools that amplify your impact and a culture that backs your ambition, you won’t just contribute, you’ll make things happen, fast.

ZoomInfo’s Core Data team builds the company and contact data behind the Go-To-Market Intelligence Platform: hundreds of millions of records, resolved to the right entity, kept accurate, and delivered to more than 35,000 customers.

The way that data is produced is changing. For years it came from deterministic pipelines: many sources, weighted and decayed, one winner per attribute. Core Data is now moving to AI-native data management, where an agent pursues the evidence for an entity, decides every attribute over all of it, and records why. The company and contact agents already exist as working, measured prototypes. What does not exist yet is the machinery that runs them at production volume, proves every version before it ships, and lets researchers and product managers operate them without an engineer in the loop.

We are hiring a Senior Applied AI Engineer to build and own that machinery. You will complete the entity-agent pipeline, build the evidence adapters behind a single envelope, run the benchmark and golden-gate system on a schedule, build the review and batch tooling the team operates, and carry the agents onto ZoomInfo’s Agentic Platform. You will work with the Principal AI Engineer who owns the benchmark harness and productionization, the product manager who owns evaluation and policy, and the company and person engineering teams that own the pipelines the agents feed. You will not be the only person who can do this work; you will be the second, so that neither of you is a single point of failure.

This is a role for someone who has shipped LLM systems in production with evaluations, builds platform for other people to operate, and can work inside a large-scale data pipeline. You should be as comfortable writing an adapter with a replayable store and a kill switch as you are reading a miss analysis with a product manager, and you should like that the prototype you take to production was built by a PM who commits code.

Run the benchmark and golden-gate system. Move judge runs, golden-gate results, arms scorecards, and cost per record from a laptop to a scheduled system against staging. Every agent version is gated automatically, with variance and drift tracked per release. The evaluation product manager sets the thresholds; you make them enforceable.

Complete the entity-agent pipeline. Build the stages that do not exist yet: enricher fan-out, location sets, and hierarchy batch files. Migrate the classifiers to the shared runner with the Principal AI Engineer so they run at production scale. Measure throughput and unit cost per stage on named cohorts, so the cost model comes from instrumentation rather than estimates.

Build the evidence adapters. Registries, live web, mail-tenant signals, and external research vendors all sit behind one envelope with a replayable vendor store, measured lift per shape, a per-vendor kill switch, and at least two sources per shape. Vendors are swappable by design.

Deploy onto the ZoomInfo Agentic Platform. Carry the company and contact agents through the production plan onto ZAP, build the ingest adapter side (envelope to match service and bulk intake), and keep a rollback path per version. Keep the DOH implementation and the platform implementation upgraded in lockstep, enforced by tooling.

Make the loop self-service. Build the review queue, tagging tools, blast-radius viewer, and batch runners that researchers and product managers use without an engineer. Measure review yield and golden-record production.

Keep the agent surfaces alive. Own the health and deployment of the company classifier, contact classifier, agents hub, and judge platform inside our internal tooling, so the team’s iteration cycle never waits on one person.

Instrument everything. Cost per call, per row, per stage; quality per version; alerts before a human notices.

• 5+ years of software engineering experience, including at least 2 years shipping LLM-based systems to production: agent loops, tool use, retrieval, structured output, and the evaluation harness that proved them
• Strong Python and TypeScript, and comfort across an API layer, a batch runner, and a small web UI for internal users
• Experience building platform others operate: adapters, runners, queues, or internal tools designed for replay, idempotency, and observability
• Working knowledge of data pipelines and entity resolution: how records move through ingestion, matching, and serving, and what a false merge costs versus a missed match
• A track record of taking a prototype to production without losing what the prototype proved, treating test sets and benchmarks as the contract
• Instrumentation by default: you know the cost and latency budget of every system you have shipped
• Strong bias for action and the tenacity to find an endpoint owner, get a decision, and unblock a runner in a fast-moving, distributed organization

Preferred

• Experience with B2B company or contact data, identity or registry data, or entity resolution at scale
• Experience with Claude or comparable frontier models, agent frameworks, and eval harnesses
• Experience deploying agents onto an internal agent platform or orchestration layer
• Background at an AI-native startup, a forward-deployed or applied AI team, or a data infrastructure company

A builder who ships. You pick the boring technology when it wins and you would rather have a running adapter than a design doc.

An evaluator by instinct. You do not release a version without a test set, a threshold, and a number for what it costs.

A platform thinker. You build things that other people run, and you design for the day the vendor goes down.

Comfortable in the data. You can read a pipeline, reason about matching, and explain why a wrong merge is worse than a missed one.

Tenacious. ZoomInfo is fast, distributed, and messy. You go find the people you need and drive the work to done.

One team. Honest, has the team’s back, celebrates wins and losses together.

You report to the Senior Manager of Product for Core Data, with the Principal AI Engineer as your technical lead, and sit inside the Core and Global Data organization under the Chief Data Officer. Your day-to-day partners are the Principal AI Engineer, the product manager who owns evaluation and data policy, the Company and Person Data product managers, the research team that grades and rules, the agentic platform team, and company and person engineering. Expect interesting problems: how do you run a judge over a million records a day for cents, not dollars? How do you make a vendor replaceable without a rewrite? How do you let a researcher run a batch and rule on the misses without paging you?

ZoomInfo (NASDAQ: GTM) is the Go-To-Market Intelligence Platform that empowers businesses to grow faster with AI-ready insights, trusted data, and advanced automation. Its solutions provide more than 35,000 companies worldwide with a complete view of their customers, making every seller their best seller.

ZoomInfo is proud to be an equal opportunity employer, hiring based on qualifications, merit, and business needs.