1. What is a Forward Deployed Engineer (FDE)?
Imagine an e-commerce company says:
"We want to build an AI chatbot for customer support."
Should you immediately start building a chatbot? No. First, understand the actual business problem.
The company might really want to reduce customer waiting time, automate repetitive questions, reduce support costs, and provide accurate answers.
So the actual requirement becomes:
Resolve customer problems faster without sacrificing accuracy or trust.
The chatbot is only one possible solution.
How an FDE thinks
A Forward Deployed Engineer works closely with customers and their teams to understand problems, build technical solutions, integrate existing systems, and take those solutions into production.
Remember: Understand the problem before choosing the technology.
2. Rule-based AI vs. Machine Learning vs. LLMs
Let's say a customer asks, "Where is my order?"
A. Rule-based AI
Developers write explicit rules.
if (message.includes("refund")) {
return "Here is our refund policy.";
}Problem: customers can express the same intent in many different ways.
- "Where is my order?"
- "My package hasn't arrived."
- "Bhai, parcel abhi tak nahi aaya."
A keyword-based bot may fail to recognize that all three concern order tracking.
B. Machine Learning
Instead of writing every rule, we provide examples.
| Input | Label |
|---|---|
| Where is my package? | Order Status |
| Cancel my purchase | Cancellation |
| Where is my refund? | Refund |
The model learns patterns from examples and predicts the category of a new message.
However, identifying a refund request doesn't actually process the refund.
C. Large Language Models (LLMs)
LLMs can interpret varied language, summarize complaints, extract information, answer questions, and help select tools.
But they can still give incorrect answers.
For example, an LLM might say a laptop can be returned within 30 days when the company's actual policy allows only seven days.
Key lesson: A convincing answer is not necessarily a correct answer.
| Approach | How it works | Main limitation |
|---|---|---|
| Rule-based AI | Humans write rules | Difficult to cover every variation |
| Machine Learning | Learns patterns from examples | Depends on data and the task it was trained for |
| LLM | Generates context-sensitive responses | Can produce incorrect or unsupported answers |
3. Why isn't an LLM alone enough?
A general LLM doesn't automatically know your company's latest policies, live order status, customer permissions, or payment records.
The application must provide the necessary information.
A grounded response is based on relevant information supplied to the model, rather than relying only on its general learned knowledge.
There is another important distinction:
- Saying "Your order has been cancelled" does not cancel the order.
- The backend must validate the request, call the cancellation API, check the result, and then confirm success.
This is why real AI applications need APIs, business rules, authorization, error handling, and monitoring.
4. AI, Machine Learning, and Deep Learning
These concepts are related:
AI is much broader than LLMs. It also includes recommendation systems, image recognition, speech recognition, and many other technologies.
5. Training vs. Inference
This is one of the most important concepts to remember.
| Stage | What happens |
|---|---|
| Training | The model learns patterns by adjusting its parameters using training data and a learning objective. |
| Inference | The trained model processes new input and produces an output. |
When your Node.js application sends a prompt to an LLM API and receives an answer, it is performing inference.
A normal prompt does not automatically retrain the model.
Memory trick: Training = learning. Inference = using what was learned.
6. From neural networks to Transformers
Older machine-learning approaches often required humans to design useful features. Neural networks can learn internal representations from data.
A neural network contains connected units arranged in layers. Deep learning uses networks with multiple layers.
Language modeling then evolved through several approaches:
- Statistical language models: Counted which words appeared together; limited context.
- RNNs: Processed text sequentially, carrying an evolving internal state.
- Attention: Helped the model focus on relevant parts of the context.
- Transformers: Made attention central to the architecture behind modern LLMs.
What is attention?
Consider:
"The animal did not cross the street because it was tired."
To interpret it, the model needs to use the surrounding context. Attention allows tokens to combine information from relevant parts of the sequence.
A Transformer contains multiple layers that refine token representations, using attention and other neural-network operations.
Remember: Attention helps connect context. Transformers use attention as a key part of their architecture.
7. What does GPT stand for?
| Letter | Word | Meaning |
|---|---|---|
| G | Generative | Creates new output, such as text, answers, and code. |
| P | Pre-trained | Trained on large datasets before being adapted to particular applications. |
| T | Transformer | Uses a neural-network architecture built around attention. |
GPT-style models generate text by predicting the next token repeatedly.
Input: The capital of India is
Predict: Delhi
Append: The capital of India is Delhi
Predict again: .
Append: The capital of India is Delhi.A token is a piece of text, such as a word, part of a word, or punctuation.
This repeated process is called autoregressive generation.
8. Interview revision
"What is an FDE?"
An engineer who works closely with customers to solve real problems and deliver production solutions.
"Why isn't an LLM enough?"
It needs relevant data, backend integrations, validation, and security.
"What is machine learning?"
Learning patterns from examples.
"Training vs. inference?"
Training adjusts parameters; inference uses the trained model.
"What is deep learning?"
Machine learning using multilayer neural networks.
"What does attention do?"
Helps a model combine relevant contextual information.
"What is a Transformer?"
A neural-network architecture that uses attention as a central mechanism.
"What does GPT stand for?"
Generative Pre-trained Transformer.
9. The memory map
If you remember only one diagram from this topic, remember this:
The history behind modern LLMs follows a similar progression: explicit rules → machine learning → deep learning → attention → Transformers → GPT.
The key takeaway: An FDE doesn't just connect an LLM to an application. They understand the business problem and build the complete system needed to solve it reliably.