.svg)
.svg)
For Instructor
Have you been following job postings at leading AI research institutes like Anthropic, OpenAI, Cohere, or at innovative enterprise AI companies? If so, you will likely have observed a common pattern. The one job title that is rapidly increasing in demand is: “Forward Deployed Engineer” (FDE).
So what does it mean to become an FDE in a time when AI can generate outputs (e.g. text; images) independently?
Traditionally, an FDE served as an intermediary between a company’s internal engineering team and the user customer — providing a customized version of a base software product to meet particular enterprise requirements. However, in the context of this new generation of AI technologies, the FDE role has evolved to be much more versatile. Rather than merely adjusting a dashboard interface, you are taking the power of AI (i.e., through foundational models like Claude or GPT-4) and utilizing it to connect the many intricate and unstructured components of real-world business processes or workflows together. Consequently, you serve as a conduit that enables companies to capture ROI from the remarkable potential of AI.
In order to succeed as an AI Forward Deployed Engineer, it is necessary to have much more than just the ability to write an engaging prompt. You must also possess three key elements that make up the "core pillars" of your role: (1) product intuition, (2) experience in coding software products and services, (3) experience interacting with customers. These three elements are essential to your success in this position. Here's how you can develop these skills and be an outstanding AI Forward Deployed Engineer.
1. Evaluating AI Agent Performance (Reality Check)
When developing a traditional application, you can be confident of the operation of your program. When coding an addition operation on two integers, you know that no matter how many times you call this addition function, it will return the same sum each time you invoke the addition function.
With AI agents, the results of AI operations are often probabilistic; they are not deterministic. For example, when working with an AI agent to summarize a legal contract, you could potentially receive slightly different or very different output each time you request the summary of the same contract.
If you're a developer, how can you assess the quality of an AI agent? How can you assure your client that the bot you're creating for him/her will not hallucinate a refund policy that is fake?
By writing evaluations.
Writing thorough evaluations is what distinguishes a typical AI hobbyist from a professional Forward Deployed Engineer. Evals are basically algorithmic tests that assess the output of your AI's operations.
To become proficient in this area, you must stop relying on your "gut instincts" (reading several output statements and thinking, "I feel like that is right") and implement structural measurement systems. Here's what an effective evaluation would include:
Producing "golden" datasets: Collaborating with domain specialists to create a grouping of ideal, verified inputs and outputs.
LLM-as-a-Judge: You'll frequently use a larger, smarter model (like Claude 3.5 Sonnet) to grade the outputs of a faster, cheaper model (like Haiku) based on strict rubrics.
Measuring Metrics: You should not only measure accuracy but also latency, token use, cost per run, and adherence to certain format specifications (like strict json).
As an FDE, you must embrace the process of measuring, adjusting, and re-measuring. The evaluations you perform provide you with a safety net that allows you to verify that an agent will perform properly in a production environment when you deploy it.
2. Creating Effective Agents for Various Workflows
Creating a simple AI chatbot to provide responses to PDF-based inquiries can be a fun weekend project. However, as an FDE, you'll be responsible for developing fully-fledged AI agents - these are autonomous systems capable of performing their own tasks based on logic, discovering resources, and executing activities independent of the user.
Today, customers do not simply wish to speak with their data; they require the AI agent to take over most of the service workload. For example, a customer sends a support ticket, your agent will extract information from the database to confirm where a delivery has been sent, reference your company policy, and generate an email to provide a solution.
In order to create agents rated at this level, you will need to understand the fundamentals of designing agent-based workflows.
Agent Interaction with External Resources: All agents require appendages, so you must have a very strong comfort leveraging the external APIs available to your models. Be it using the APIs to pull down live weather data, posting changes to your code repository in GitHub, or updating the customer relationship management system via Salesforce.
Workflow Coordination and Routing: Generally speaking, a single request from a client will not adequately meet the needs of complex enterprise requests, therefore a system must exist for routing the initial request from an agent and passing it on to multiple sub agents for processing based on the nature of the request.
The System of Tracking State and Memory: An agent requires a reliable memory of previous actions to guide future actions, specifically when executing multi-step processes across many steps.
Creating Boundaries (Guardrails): In the event that the agent does not possess a definitive response, its boundaries provide parameters so that it fails amicably instead of providing a false affirmative.
The Evolution of the Chat Interface to an Agent: The premier FDEs are examining all current internal tools possessed by the organization and attempting to construct agents that perform these functions with the same level of proficiency as humans.
3. Acquiring a Solid Understanding of the Customer's Challenges from A to Z
This is where the value comes from, and is the key factor in getting hired by companies like Anthropic.
You may be an exceptionally talented developer, capable of creating the most sophisticated orchestration systems that involve multiple agents; however, if you do not fully comprehend the commercial challenges that your customers face, you will likely create a fantastic but ineffective machine.
The customer success stories at Anthropic's website also provide insight into this. Customer successes have very little to do with the intelligence of their models, but more to do with how well the AI system is integrated into the workflow of the person who will be using it.
As an FDE, your first responsibility is not to code, it is to investigate.
For example, if a customer states they would like an AI system to help them with legal review automation, you need to find out more.
: What exactly do lawyers seek to gain from the documents they're reading?
: Where does this information exist initially before they read it?
: In what format do they need the final product produced?
If the AI misses some terms, consequences could be destructive, killing the company & eliminating trust in the vendor of AI software or hardware products.
The flow of a contract from the moment it enters a company until a person makes a decision is done in this manner; The brand will not understand their own flow until you make that information available to them.
After understanding how the flow is created "from beginning to end", one can find the bottlenecks in the process that would benefit from an AI agent the most. Therefore, you will have a different relationship with the customer as a trusted partner rather than someone who just creates code. You will assist users by removing confusion and human logic from their businesses, providing pure evidence of AI systems able to evaluate & process the data accurately.
