AI Platform Engineer

About the company

Forterra builds autonomous and interoperable mission systems that automate vehicles and communications for defense and logistics operations.

About this role

About the role Forterra runs its engineering on agentic systems, not chatbots. Requirements, safety analysis, test generation, code review, and validation are increasingly performed by fleets of agents operating against Forterra's real systems of record. This role builds and owns that substrate: the agent definitions, skills, MCP servers, guardrails, and evaluation harnesses that determine whether hundreds of engineers get leverage or just burn tokens. The hard part is not wiring an API. It is judgment, such as knowing which problems should be handed to an agent, which should stay deterministic, when to use a classical data analysis method, and how to prove an agent's output is trustworthy. You will work across the entire company: systems engineering, safety, software, product, IT, cybersecurity, and executive level leadership. Almost none of it sits inside your reporting line. Influence and growth directly come from performance and earned credibility. You will have direct visibility to chief executive leadership. What you'll do Design agentic systems, not prompts Author agent definitions and context as engineered artifacts, with explicit role separation, tool allowlists, model selection, and hard behavioral limits. Design multi-agent architectures where roles are adversarial and reliable by construction. Build skills with disciplined triggering, progressive disclosure, and non-overlapping ownership boundaries - so capability scales without the context window collapsing. Orchestrate work across multiple model vendors, harnesses, and agent runtimes. Make and defend the call on whether a problem warrants an agent at all. Know when a deterministic script or webhook is the correct answer, and how to engineer away from an agent left polling in a loop. Make agents part of the engineering infrastructure Design how agents get invoked across the full range of trigger patterns — a developer working interactively at a terminal, a nightly scheduled run, an event-driven trigger like CI, and long-running unattended work. Know the failure modes for each, and ship the result as shared, versioned engineering infrastructure rather than one-off scripts. Define the autonomy gradient explicitly: which agents are permitted to act, and which may only propose. Engineer unattended agents for the absence of a human: bounded scope, idempotent reruns, deliberate fail-open versus fail-closed decisions, cost and runtime caps, structured verdicts, and ephemeral scoped credentials that keep them auditable. Design the human-facing side of autonomous work: how an agent escalates when it is genuinely blocked, how and when it asks for a human in the loop, and how a fleet of concurrent agent sessions stays observable to the humans accountable for them. Prove agent output is trustworthy Design defensible evaluations of AI tooling: controlled A/B trials, ablation studies, blinded LLM judges, contamination and leakage controls, reported uncertainty, and honest null results. Build validation harnesses that distinguish genuine signal from hollow green. Instrument your own systems against complexity drift: measure whether each added skill, agent, or tool is actually pulling its weight in quality, cost, and turns. Build the guardrails Design structural controls for agents operating near controlled information: layered redaction, egress filtering, enforcement at the data layer rather than the prompt layer. Anticipate cross-boundary failure modes unique to agents. Connect agents to the enterprise Build and operate the connective tissue between agent tooling and Forterra's infrastructure & processes: requirements management, issue tracking, documentation, source control, diagramming, and unstructured data. Own enterprise identity integration end-to-end, including federated and government-cloud authentication, and specify precisely what changes IT needs to make. Run these as real production services with real uptime expectations, including incident response and root-cause write-ups. Drive adoption Turn tribal knowledge into onboarding paths, curricula, setup automation, and self-diagnosing tooling that lets an engineer's own agent fix its broken configuration. Teach transferable technique rather than button-clicking, including curriculum for non-software audiences — systems engineering, safety, program management, design, business operations — and design yourself out of delivery so training scales past you. Maintain the platform other engineers contribute to: review their work, set and hold the standards for acceptable agent patterns, grow the contributor base, and convert user pain into prioritized organizational demand. Minimum Qualifications 3+ years building software systems, with meaningful recent depth in LLM-based or agentic systems Demonstrated fluency with modern agent architecture: tool/function calling, MCP or an equivalent tool protocol, context management, retrieval, multi-agent orchestration, and hooks or middleware for policy enforcement Practical judgment about model selection and cost/latency/quality tradeoffs, and a habit of tracking the frontier rather than relying on last year's assumptions Experience evaluating AI system quality with rigor Experience with analysis of unstructured data Experience running non-interactive workloads in production: CI pipelines, scheduled jobs, or event-driven triggers Strong Python intuition plus comfort in shell, CI, and Linux service operation Experience building and maintaining production applications Working understanding of enterprise identity and authorization (OAuth2/OIDC/SAML, tokens, scopes, least privilege) Excellent written AND oral communication Preferred Qualifications Strong background in statistics and/or classic machine learning AND deep learning Experience fine tuning LLMs on codebases and enterprise data (DPO, QLoRA, GRPO) Contributions to open source agent harnesses Knowledge-graph, ontology, or retrieval architecture work across heterogeneous unstructured enterpris