RAD HIRES

Full-Stack AI Engineer

Remote, LATAM. Full-time placement with a US-based tech startup. Mid-to-senior.

About the role

Rad Hires places senior LATAM engineers into US tech startups that need someone who can take an ambiguous problem from blank page to shipped MVP, with AI integrated at every stage.

We are hiring architects and harness builders . The model is the engine. The harness is everything around it: repo structure, MCP servers, tools, sub-agents, evals, guardrails, internal UIs, documentation, traces, and automation pipelines. A good harness makes a mediocre model useful in production. A bad harness wastes a great one. We hire the people who build the harness.

This is explicitly a mid-to-senior role. AI tooling amplifies the engineering instincts you already have. It does not install them where they are missing. We are looking for engineers whose judgment about systems, maintainability, and tradeoffs was earned before the model could help.

Who you will work with

Our clients are US-based tech startups, typically seed through Series B, building AI-forward products. Small teams, fast cycles, real customers from day one. The kind of place where one engineer can ship a feature from blank page to first user without three meetings to get permission. You will work directly with the founding team, product, and operations, not behind layers of management.

What you will do

• Design codebases so AI agents can work in them with high success rates. Repo conventions, AGENTS.md or CLAUDE.md hierarchies, machine-readable structure, and the guardrails (tests, linters, hooks) that keep agents honest while they edit.

• Build MCP servers, custom tools, and focused sub-agents that wrap business logic and expose it to both AI clients and non-technical teammates.

• Set up retrieval, schemas, and context systems so models do not have to guess.

• Build internal tools (admin panels, CMS layers, visual editors, configuration UIs) that let marketing, ops, and support make routine changes safely, without filing tickets.

• Design evals, define autonomy boundaries (always-do, ask-first, never-do), and own the observability that keeps agent-driven systems honest in production.

• Take features from blank page to first-customer iteration, across the stack, orchestrating AI tooling deliberately to compress the build cycle without losing the plot.

What we look for

• Engineering judgment, first. You can decompose an ambiguous business problem into a system: data, interfaces, control flow, failure modes, who owns what. You reason cleanly about tradeoffs between deterministic code, agent loops, retrieval, and human-in-the-loop. You surface non-functional requirements (latency, cost, blast radius, recoverability) without being asked.

• Active harness experience. You have built and shipped at least one MCP server, custom tool surface, or production agent workflow. You can defend the tool boundaries you chose and what changed after real usage. You have opinions on hooks, sub-agents, and orchestration frameworks (LangGraph, CrewAI, AutoGen, or custom), and you are not married to any one stack.

• Static harness instincts. You have a point of view on repo layout, naming, and module boundaries that make context windows tractable. You maintain honest AGENTS.md or CLAUDE.md files and treat agent failure as usually an environment-and-context problem, not a reasoning problem.

• Safety harness discipline. You treat evals as first-class engineering work: offline sets, regression gates in CI, LLM-as-judge with human calibration. You have set up traces, tool-call logs, cost and latency budgets, and dashboards a non-engineer can read. You plan for model outages, regressions, and cost spikes.

• End-to-end product velocity. You have taken a product or feature from blank page to a customer using it: data model, backend, frontend, deploy, monitoring, basic ops. You do not need permission to touch any layer. You know when an MVP is real versus when it is a demo that will not survive contact with reality.

• AI enablement for non-engineers. You have shipped internal tools that non-engineers depend on, with preview, staging, approvals, and audit log designed in from the start, not bolted on. You believe the goal is to make non-engineers more capable, not more dependent on engineering.

• Verification over delegation. You treat AI-generated code as a draft, never as authoritative. You can walk us through a specific time you rejected or rewrote AI-generated code: what you spotted (hallucinated APIs, silent error swallowing, scope creep from the agent), and how you caught it.

• Engineering fundamentals (hard floor). You have shipped real production systems and can talk about what you learned operating them over time. Comfortable in TypeScript or Python plus at least one backend stack. You can model data in relational and document stores, understand HTTP, auth, and security basics, and have shipped with containers, CI, and either cloud or self-hosted infrastructure.

• Communication. You can describe an architectural decision in plain language and in technical language, and pick the right register for the audience. You write clearly in chat, docs, and commits.

Signals we love

• A portfolio of small, weird, useful internal tools, not just big projects.

• A multi-agent or orchestrated system in production with explicit role boundaries, segmentation, and a story about what broke first.

• Examples of helping non-engineers on past teams do more of their own work.

• Strong opinions on hooks, slash commands, custom skills, and sub-agents, and when each is the right primitive.

• Tried at least one new AI tool in the last sixty days. Can name two or three things that meaningfully changed in the field this quarter.

This role is not for you if

• Your value pitch is "I write code fast."

• You ship AI-generated code without reading it, testing it, or being able to defend it.

• You have never built anything a non-developer actually uses.

• You build "god mode" agents with broad toolsets and no clear role boundaries.

• You treat AI as autopilot, or as an existential threat. Neither posture works here.

• You have not touched a new AI tool in months.

Logistics

• Location: Remote, anywhere in LATAM.

• Hours: Minimum 4 hours of daily overlap with US Pacific or US Eastern business hours.

• Language: Professional English fluency required, written and spoken. You will be in daily contact with US-based product, design, and operations teammates.

• Engagement: Full-time placement with a US-based tech startup client. Rad Hires sources, vets, and matches. You work directly with the client team.

• Compensation: Competitive hourly rate in USD based on experience and technical evaluation.

How to apply

We hire on evidence, not credentials. Send us:

• A short note (skip the cover letter) on the most interesting harness you have built or repo you have made AI-legible.

• Links to real artifacts: repos, MCP servers, admin UIs, eval suites, dashboards, traces. Show, do not tell.

• Your CV.
  

Apply here: recruiterflow.com/radhires/jobs/369

What happens next

We respond to every application within five business days, either way.