Founding Platform Engineer, Agent Runtime

Zingage•New York City, NY

9h•Onsite

About The Position

Home care is a $130B industry that runs on the telephone, and we are the AI that answers it. Zingage's agents coordinate scheduling, on-call, and intake for 120+ home care agencies - 300K+ patient visits a month ride on our decisions, across every major EMR in the industry. We recently closed our Series A, backed by Bessemer, Bertelsmann Investments, and Yosemite (Reed Jobs), and the team is deliberately small and dense - operators and engineers from Ramp, Uber, Tennr, Datadog, and Verkada. Headquartered in New York. The problem: Our agents act in the physical world: they fill shifts, move visits, and write to the system of record that governs care for frail patients. An agent is only as good as the runtime beneath it - and the runtime for production agents in a regulated industry does not exist yet. Building it means solving four problems your best colleagues would agree are open: 1. Distributed systems where the component is the source of nondeterminism. Forty years of systems design assumes deterministic components in an unreliable world. An agent runtime inverts the assumption: the actor itself is stochastic. What does idempotency mean for an agent action? What is a transaction when one participant is an LLM and the other is a twenty-year-old EMR with no locking semantics? 2. Freshness as a per-decision SLO. "How stale is too stale to act on?" has a different answer for a 2AM call-off than for next week's schedule change. We need staleness budgets tied to real-world risk, enforced against four EMRs with no webhooks, brutal rate limits, undocumented semantics, and silent failure modes - CAP theorem where the partition is permanent and political. These upstreams were never designed to be built on. That is not the obstacle; it is why this layer, once built, is the moat. 3. Replay of a stochastic actor. You cannot replay the model, so you must replay the world: capture every agent trace - inputs, tool calls, upstream state - completely enough that a new model can be interviewed against two years of production reality overnight. 4. Time-travel debugging for agents. The trace corpus this produces is the most valuable asset the company will ever own, because it is how we adopt every new model first while competitors spend a quarter stabilizing. 5. Safety as unrepresentability. A production outage taught us the principle the hard way: no prompt can make the model emit a field the schema forbids. Your job is to generalize it - a capability and governance system where unsafe actions against a patient's record are not discouraged or detected but unrepresentable at the boundary between a stochastic planner and a real-world effector. Zero unauthorized writes, audit-verified, forever, in a HIPAA-regulated industry where the audit trail must satisfy payers and regulators, not just engineers.

Requirements

Technical judgment
Ownership mindset
Role clarity
Experience with distributed systems
Understanding of nondeterminism in systems
Experience with LLMs and their integration into systems
Familiarity with EMR systems and their limitations
Experience with SLOs and their enforcement
Understanding of CAP theorem and its implications
Experience with data replay and time-travel debugging
Knowledge of safety and security principles in regulated industries (HIPAA)
Ability to design and reason through system trade-offs
Experience with blameless postmortems
Experience shipping risky changes behind feature flags with named rollback owners
Commitment to a 60-day durability standard for fixes

Nice To Haves

Experience with systems where actors are stochastic
Familiarity with EMRs lacking webhooks, with brutal rate limits, undocumented semantics, and silent failure modes
Experience building on top of systems not designed for integration

Responsibilities

Founding ownership of an entire layer, not a slice of someone's data platform - the whole runtime, all four planes, the architecture decisions, and in time the team that grows around it.
Build the first version of each plane, then hire the team that owns them.
Generalize safety as unrepresentability: a capability and governance system where unsafe actions against a patient's record are not discouraged or detected but unrepresentable at the boundary between a stochastic planner and a real-world effector.
Ensure zero unauthorized writes, audit-verified, forever, in a HIPAA-regulated industry where the audit trail must satisfy payers and regulators, not just engineers.
Drive the screen when an EMR has no API, acting as the universal adapter that ends our dependence on upstream permission.