Applied AI Engineer, Agent Systems

The Electric Plant•San Francisco, CA

1d•Onsite

About The Position

The Electric Plant Co. is building a new category of plant and tree intelligence. Our IoT hardware measures the electrical signals in plants and trees, paired with environmental data, and our foundation model, Bombadil, decodes those signals into real-time insights about plant health, growth, and stress. We're a small, fast-moving company in San Francisco working at the intersection of biology, hardware, AI, and IoT. Bombadil can decode a plant’s physiological responses to stress — the layer that turns that into something a grower acts on doesn't exist yet, and that's what this role builds. The Role 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.

Requirements

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.
Strong plus Integration grit against client and legacy systems: enterprise APIs, CSV over SFTP, email deliverability.

Responsibilities

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.