The role
You are the person who makes AI work inside a client’s actual environment, with their real data, their legacy systems, their compliance constraints, and their skeptical operations manager who has seen three failed software rollouts already.
The gap between a demo that impresses a CFO and a system that runs unattended on Monday morning is enormous, and this role exists to own that gap. You’ll embed with a client for weeks at a time, build the thing, get it into production, and then train their people to run it without you.
A concrete example of the work: one of our engagements is an AI-assisted workflow that drafts sales recap emails with a pricing recommendation attached. The design that shipped puts the pricing math and plan-threshold rules in deterministic code, uses Claude to extract signals from call transcripts and narrate the recommendation, screens every output fail-closed (versus fail-open), and gates the send behind an explicit human approval. If that division of labor strikes you as correct, where code decides, the model writes prose, and a human authorizes the irreversible action, you’ll recognize how we’re building for our clients.
What you’ll do
- Sit with client teams to find the workflows where an agent actually pays for itself, and say so plainly when one doesn’t
- Work with a client to design and ship production agent systems using the Claude Agent SDK, deployed on Vertex AI within the client’s own Google Cloud project
- Split every system into its deterministic and model-based parts before writing a line of it: rules and math live in code, judgment and prose live in the model, and irreversible actions live behind a human gate
- Integrate agents with the systems clients already run: Google Workspace, BigQuery, CRMs, ticketing systems, internal APIs
- Build evaluation suites alongside the first version, not after it. Create golden datasets from real client history, code-graded checks for rules and math, and model-graded checks for quality, so we can prove a system works before it touches production, and catch it when it drifts afterward
- Own the unglamorous middle: auth, error handling, retrieval scoping, cost controls, logging, permissions, and rate limits, with support from your HiView colleagues and client team resources
- Hand off properly, documentation, runbooks, escalation paths, and live training so the client’s team owns the system when you leave
- Deliver hands-on enablement sessions from our training catalog: advanced Claude Code workflows, Skills, plugins and versioning, and agent pattern selection, each with a takeaway artifact the client keeps
- Feed what you learn back into our frameworks and company playbooks
What we’re looking for
- 4+ years writing production software in Python or TypeScript, with real ownership of things that ran and broke and got fixed
- Hands-on experience building LLM-backed systems that went to production, not pilots, not notebooks
- Working knowledge of the Anthropic stack: the Claude Agent SDK, tool use, MCP, prompt caching, and structured outputs
- Interest in developing your own skillset with expanded Claude and Google Cloud certifications, time will be allocated in your role for this
- The judgment to tell a client that the thing they asked for is the wrong thing to build
- Solid cloud fundamentals; working knowledge of GCP accepted, GCP pros will be prioritized
How to apply
Send a resume and links to anything you’ve built with agents. If you have a repo using the Agent SDK or an MCP server you’ve shipped, lead with that, it tells us more than a cover letter will.
Email your application to [email protected].