You'll build the application layer that sits between our foundation model and the people who use it: the client's context layer, insight and policy cards, learning loops, agents, and scheduled work. You'll start embedded in the field with our Deployment Lead on our first real customer, shipping things a grower actually uses to run their orchard. From month six onward, you'll convert what worked into the reusable protocols and tools that make every deployment after the first faster to stand up.
This is a senior, individually-contributing role with no direct reports. You'll report to our Director of Engineering, with a direct line to our CEO on what the system could do next.
You'll work closely with the engineer who owns our deterministic data plane (ingest, fleet, telemetry), the team that owns our foundation model, and the person who owns the client relationship and translates what growers need into what you build.
Client context layer and retrieval. Built and measured, not just assumed to work.
The tool schema exposing our platform's data and insights, including uncertainty in the return payload.
The product's agent systems — skill and instruction files as versioned, testable, reviewable artifacts, and the scheduled/event-driven work built on them, durable, idempotent, and observable.
One insight pipeline rendering to chat, email, PDF, and API.
The context schema and card structure that make Platform's multi-tenant isolation and client IP protections possible — including how clients' data and the insights engineered from it are protected.
Actuation. When the system carries a client-set policy into their systems — an irrigation setpoint, a crew instruction - you own making sure it executes reliably every time: safe to re-run, fully trackable, and only after the approvals and deployment steps the client has set up..
Eval suites, trace capture, drift monitoring, and per-client cost accounting.
Qualification
Has built durableHas built eval suites
Required
Has built and measured production AI retrieval and context assembly, with quantification of the system’s usefulness, not demo impressions.
Consistently pushes repeatable work into deterministic code and spends model calls only where judgment is genuinely required.
Has shipped multi-tenant isolation in production: tenant-scoped data and indexes, per-client credential scoping, deletion that actually deletes.
Has built a config or skill system that a non-engineer colleague used daily and modified without you.
Has operated an agent system in production, not a prototype, with a real story about silent drift, a cost blowup, or a cron that double-fired.
Has built durable, idempotent, observable scheduled or event-driven agent work.
Has built eval suites, trace capture, or drift monitoring, and can speak to where ground truth came from and how often it arrived.
Likely has a full-stack or front-end web/application engineering background who made a deliberate turn toward agentic systems, rather than a long-tenured backend or data engineer moving into this space for the first time. Total years matter less than this shape.
Integration grit against client and legacy systems: enterprise APIs, CSV over SFTP, email deliverability.
This isn't a role for someone who wants a settled spec and a quiet backlog. Requirements will arrive from an orchard, mid-season, and be wrong. It's also not a research role — you'll consume and expose what the model does, not train it — and not an infrastructure role, since the data plane underneath you is someone else's deterministic domain. If you're excited by the orchestration toolkit itself more than by what it takes to get something working in a grower's hands, this isn't the fit. If you want to build the layer that makes a decoded plant signal into something a business actually acts on, and to do it first in the field and then in code that outlasts the first customer, we'd like to talk.
