Forward deployed engineer
Take AI into a customer’s world: their data, their APIs, their rules.
At the end: Turn a customer’s problem into a working AI system on their stack, and prove it works with evals.
Stage 1: Foundations
The tools you reach for on day one at a customer.
- Python, properly: Fast, tested Python for glue code and services.
- SQL from zero: Answer questions from the customer’s data yourself.
- Linux on call: Find your way around a server you did not set up.
- Git without fear: Work in someone else’s repository without fear.
Stage 2: Ship services
Their API, your service, in their cloud.
- APIs with FastAPI: An API the customer’s systems can call.
- Docker, start to ship: Ship it the same way to every environment.
Stage 3: AI on their data
The core of the job.
- Building with LLMs: The raw model API, costs and limits.
- RAG in Python: Answers from their documents, with citations.
- Agents from scratch: Agents that act on their systems, within limits.
- MCP servers and clients: Plug their tools into any assistant.
- Evals that catch regressions: Show the customer it works, in numbers.
- Harness engineering: Make the agent reliable in their environment.
- Claude and the Anthropic SDK: Claude on their documents: tools, caching, vision and PDFs.
- Build an AI gateway: Keep their AI bill and uptime in check.
- AI security: attack and defend: Answer their security team before they ask.
- Build a voice agent: When the customer’s channel is the phone.
Stage 4: Customer projects
The kind of build you hand over.
- AI support copilot project: A support copilot on their policy.
- AI invoice reader project: Their invoices into their database.
- MCP server project: Their tools behind an MCP server.
- AI ordering agent project: An agent that acts on their systems, with approval.