Staff Engineer, Agent Systems

CargoSprint

36d•Hybrid

About The Position

Most companies are building AI demos. We are building agent infrastructure. CargoSprint is deploying production agents across our internal operations — sales, finance, customer ops, engineering productivity — and we need a Staff Engineer who can design the systems that make those agents reliable, observable, and genuinely useful. You will own the agent systems layer: the retrieval pipelines that ground our agents in real organizational knowledge, the orchestration architecture that makes multi-step workflows predictable, the tooling and APIs that connect agents to the systems our teams run on, and the observability layer that tells us when something is degrading before a user notices.

Requirements

Extreme ownership — you care about what happens to the systems you build after they ship. Degradation, drift, and silent failures are personal.
8+ years of engineering experience, with meaningful time spent building systems that run reliably under real production load
A track record of technical decisions you made, owned, and lived with — including the ones that turned out to be wrong and what you did about them
Strong business judgment — you understand that a technically elegant agent nobody uses is a failure. You can read a workflow, identify the real cost, and design for adoption, not just correctness.
Excellent communication in English — you can explain a retrieval architecture to a product manager and a vector indexing strategy to a staff engineer, and you know which explanation to give in which room
Willingness to travel to CargoSprint's Guadalajara, Mexico office as needed — the workflows you are designing systems for live there, and understanding them firsthand matters
Expert-level Python — idiomatic, well-tested, production-grade. You write code that the next engineer can understand and extend.
Deep RAG system design experience — you have designed and operated retrieval pipelines in production: chunking strategies, embedding model selection, hybrid search, re-ranking, context window management, and retrieval evaluation. You know the failure modes intimately.
Agent orchestration architecture — LangGraph, LangChain, or equivalent; you have designed multi-step agentic workflows with tools, memory, branching logic, and human-in-the-loop patterns that are predictable under real usage
LLM integration and prompt engineering — you understand how to structure prompts for reliability, how to version and evaluate them, and how to manage the gap between model capability and production behavior
Vector databases and search infrastructure — pgvector, Pinecone, Weaviate, or equivalent; you know when to use dense vs. sparse retrieval and how to build an evaluation harness to measure retrieval quality
FastAPI and backend service design — you build the infrastructure your agent systems run on with the same rigor as the systems themselves
Observability and production operations — distributed tracing, structured logging, alerting, LLM-specific evaluation pipelines; you know what good looks like before something breaks
DevOps fundamentals — Docker, Kubernetes, CI/CD; you own what you ship all the way to production

Nice To Haves

You think in systems. When someone describes a workflow problem to you, you are already modeling the retrieval architecture, the orchestration graph, the failure modes, and the observability layer — before they finish the sentence.
You have built agent systems that run in production, not just in demos. You understand what makes agentic workflows fail — context bleed, retrieval drift, non-deterministic tool calls, silent degradation — and you design defensively against all of it. You have strong opinions about agent architecture because you have been burned by the wrong decisions and learned from them.
You are passionate about solving complex problems and believe in lifelong learning, constantly staying on the cutting edge of what is possible with LLMs, retrieval, and agent orchestration. And you want to do it somewhere that will actually put your work into production and measure whether it worked.
Familiarity with Temporal
Familiarity with OpenAI, Claude
Familiarity with Pulumi

Responsibilities

Design and own the agent systems architecture — retrieval, orchestration, tool integration, and evaluation — as a coherent, production-grade platform
Build RAG pipelines that ground agents in real CargoSprint data: indexing strategies, chunking, embedding models, retrieval evaluation, and freshness maintenance
Design orchestration patterns for multi-step agentic workflows using LangGraph or equivalent — with explicit attention to failure modes, non-determinism, and graceful degradation
Build and maintain the tool and integration layer that connects agents to production systems — Salesforce, HubSpot, Postgres, internal APIs — with the error handling and retry logic that production demands
Instrument everything: distributed tracing, latency dashboards, retrieval quality metrics, LLM output evaluation pipelines
Establish reusable agent primitives and internal engineering patterns so the team builds the next agent faster and more reliably than the last one
Partner with the engineers building individual agents to review architectures, catch design mistakes early, and raise the overall quality bar
Travel to CargoSprint's Guadalajara office as needed to work directly with the operational teams whose workflows the agents are being built around
Use AI coding tools to accelerate your own development and set the standard for how the team works with them