N
Neha Sharma
Guest
If youโre asking yourself, โWhere do I even start in AI engineering?โ โ youโre not alone. The landscape is crowded with buzzwords (AI, ML, DL, GenAI, MLOps), and itโs easy to feel lost. A few months ago, I had the same question: Where should I begin โ AI, ML, or infrastructure?
So, I went back to step zero to get clarity. In todayโs blog, weโll unpack the differences between AI and ML, and how to decide where to start.
Youโve probably seen the classic diagram that looks like nested circles. To understand from where to start in AI domain understanding this diagram will help you a lot. Letโs put some meaning behind it:
1. AI (Artificial Intelligence)
From a 10,000-foot view, AI is about building systems that mimic human-like intelligence. This includes everything from rule based expert systems companies used to todayโs generative AI models.
2. ML (Machine Learning)
A subset of AI. Instead of writing explicit rules, ML systems learn from data to make predictions or classifications. eg, predicting house prices based on location, size, and historical sales.
3. DL (Deep Learning)
A subset of ML that uses artificial neural networks with multiple layers. Deep learning shines when you have massive datasets and GPU/TPU hardware. eg, computer vision (image recognition), natural language processing (chatbots, LLMs), and speech recognition.
The common thread across all three? Data.
No matter which path you choose โ AI applications, ML models, or deep neural networks โ everything starts with clean, structured, and sufficient data. In short: data is the new oil.
As engineers, the question becomes: which domain should you pick โ AI, ML, or DL? To decide, letโs break it down layer by layer.
1. AI โ Application Layer
Leveraging pre-trained or foundation models (OpenAI GPT, Claude, Gemini, Llama 3, etc.) to build intelligent applications. Most GenAI startups fall into this category.
Tech stack examples: LangChain, LlamaIndex, Pinecone, FAISS, Hugging Face APIs.
2. ML/DL โ Model Building Layer
Training your own models from scratch or fine-tuning existing ones. This requires understanding algorithms, data preprocessing, feature engineering, and evaluation.
Tech stack examples: Scikit-learn, TensorFlow, PyTorch, Hugging Face Transformers, Weights & Biases.
3. MLOps โ Infra โ Scaling Layer
Making sure models donโt just work on your laptop but actually run in production at scale. Think data pipelines, CI/CD, model monitoring, GPU clusters, and distributed systems.
Tech stack examples: Kubernetes, Airflow, MLflow, Kubeflow, Amazon SageMaker, Ray, Docker.
Hereโs the practical decision-making guide:
Want to build cool apps quickly?
Start with AI (application layer). Youโll work with APIs, prompt engineering, and frameworks like LangChain to turn pre-trained models into usable apps.
Fascinated by algorithms, optimisation, and math?
Go deeper into ML/DL. Youโll learn to design neural networks, tune hyperparameters, and explore architectures like CNNs, RNNs, and Transformers.
Love DevOps, cloud, and distributed systems?
Explore MLOps / infra. This path is about ensuring reliability, scalability, and automation of AI/ML workflows.
Letโs break down the day-to-day work across each role.
You have a product idea (eg, builiding a chatbot that answers customer queries by combining OpenAI GPT with your companyโs documents (RAG pipeline).) and now as an AI engineer you will:
1 . Choosing and integrating pre-trained models (GPT, Stable Diffusion, Whisper).
2 . Building workflows around LLMs (RAG, agents, tool use).
3 . Handling prompt engineering, context windows, embeddings, and vector databases.
4 . Scaling apps to handle real-world usage.
5 . Considering hardware and latency constraints (e.g., running on GPU vs CPU, on-device inference).
You are building, training, and testing models (e.g, Train a deep learning model to classify writing into shakespear vs not-shakespear using PyTorch and fine-tuning ResNet.)
1 . Collecting, cleaning, and annotating datasets.
2 . Splitting data into train/validation/test sets.
3 . Feature engineering (normalisation, categorical encoding).
4 . Training models (RandomForest, Logistic Regression, or Transformers).
5 . Hyperparameter tuning (Grid Search, Bayesian optimisation).
6 . Evaluating with metrics (accuracy, F1, ROC-AUC, BLEU).
7 . Optimising for inference speed and resource usage.
You are an engineer who is working with DL or AI engineers. eg Build an ML training pipeline with Airflow that trains models nightly, logs metrics to MLflow, and deploys the best-performing version to Kubernetes.
1 . Automating data ingestion with ETL pipelines.
2 . Containerising models with Docker.
3 . Orchestrating workflows with Airflow or Prefect.
4 . Deploying models to Kubernetes or SageMaker endpoints.
5 . Monitoring for data drift and model degradation.
6 . Handling distributed training (Horovod, Ray).
7 . Ensuring reproducibility (MLflow, DVC, GitHub Actions).
Another important dimension is compute infrastructure:
1 . AI Applications: Often run on cloud APIs. You donโt worry much about GPUs, unless you want to self-host.
2 . ML/DL Models: Training deep networks can require powerful GPUs/TPUs (NVIDIA A100, H100, TPU v5).
3 . MLOps: Involves managing GPU clusters, storage, and networking (think: Kubernetes, Proxmox, or cloud GPU farms).
Think of it this way:
The ecosystem is huge and evolving fast. You donโt need to master everything at once. Pick the layer that excites you most, get hands-on, and then branch out as your curiosity grows.
At the end of the day, all three paths are connected โ many engineers eventually end up wearing multiple hats. Start where you are most comfortable, and build from there.
Continue reading...
So, I went back to step zero to get clarity. In todayโs blog, weโll unpack the differences between AI and ML, and how to decide where to start.
The Big Picture: AI, ML, and DL
Youโve probably seen the classic diagram that looks like nested circles. To understand from where to start in AI domain understanding this diagram will help you a lot. Letโs put some meaning behind it:
1. AI (Artificial Intelligence)
From a 10,000-foot view, AI is about building systems that mimic human-like intelligence. This includes everything from rule based expert systems companies used to todayโs generative AI models.
2. ML (Machine Learning)
A subset of AI. Instead of writing explicit rules, ML systems learn from data to make predictions or classifications. eg, predicting house prices based on location, size, and historical sales.
3. DL (Deep Learning)
A subset of ML that uses artificial neural networks with multiple layers. Deep learning shines when you have massive datasets and GPU/TPU hardware. eg, computer vision (image recognition), natural language processing (chatbots, LLMs), and speech recognition.
The common thread across all three? Data.
No matter which path you choose โ AI applications, ML models, or deep neural networks โ everything starts with clean, structured, and sufficient data. In short: data is the new oil.
The Developerโs Perspective
As engineers, the question becomes: which domain should you pick โ AI, ML, or DL? To decide, letโs break it down layer by layer.
1. AI โ Application Layer
Leveraging pre-trained or foundation models (OpenAI GPT, Claude, Gemini, Llama 3, etc.) to build intelligent applications. Most GenAI startups fall into this category.
Tech stack examples: LangChain, LlamaIndex, Pinecone, FAISS, Hugging Face APIs.
2. ML/DL โ Model Building Layer
Training your own models from scratch or fine-tuning existing ones. This requires understanding algorithms, data preprocessing, feature engineering, and evaluation.
Tech stack examples: Scikit-learn, TensorFlow, PyTorch, Hugging Face Transformers, Weights & Biases.
3. MLOps โ Infra โ Scaling Layer
Making sure models donโt just work on your laptop but actually run in production at scale. Think data pipelines, CI/CD, model monitoring, GPU clusters, and distributed systems.
Tech stack examples: Kubernetes, Airflow, MLflow, Kubeflow, Amazon SageMaker, Ray, Docker.
Choosing Your Path
Hereโs the practical decision-making guide:
Want to build cool apps quickly?
Start with AI (application layer). Youโll work with APIs, prompt engineering, and frameworks like LangChain to turn pre-trained models into usable apps.
Fascinated by algorithms, optimisation, and math?
Go deeper into ML/DL. Youโll learn to design neural networks, tune hyperparameters, and explore architectures like CNNs, RNNs, and Transformers.
Love DevOps, cloud, and distributed systems?
Explore MLOps / infra. This path is about ensuring reliability, scalability, and automation of AI/ML workflows.
How day-to-day work looks?
Letโs break down the day-to-day work across each role.
1. AI Engineering (Application Layer)
You have a product idea (eg, builiding a chatbot that answers customer queries by combining OpenAI GPT with your companyโs documents (RAG pipeline).) and now as an AI engineer you will:
1 . Choosing and integrating pre-trained models (GPT, Stable Diffusion, Whisper).
2 . Building workflows around LLMs (RAG, agents, tool use).
3 . Handling prompt engineering, context windows, embeddings, and vector databases.
4 . Scaling apps to handle real-world usage.
5 . Considering hardware and latency constraints (e.g., running on GPU vs CPU, on-device inference).
2. ML/DL Engineering (Model Building Layer)
You are building, training, and testing models (e.g, Train a deep learning model to classify writing into shakespear vs not-shakespear using PyTorch and fine-tuning ResNet.)
1 . Collecting, cleaning, and annotating datasets.
2 . Splitting data into train/validation/test sets.
3 . Feature engineering (normalisation, categorical encoding).
4 . Training models (RandomForest, Logistic Regression, or Transformers).
5 . Hyperparameter tuning (Grid Search, Bayesian optimisation).
6 . Evaluating with metrics (accuracy, F1, ROC-AUC, BLEU).
7 . Optimising for inference speed and resource usage.
3. MLOps / Infra Engineering (Scaling Layer)
You are an engineer who is working with DL or AI engineers. eg Build an ML training pipeline with Airflow that trains models nightly, logs metrics to MLflow, and deploys the best-performing version to Kubernetes.
1 . Automating data ingestion with ETL pipelines.
2 . Containerising models with Docker.
3 . Orchestrating workflows with Airflow or Prefect.
4 . Deploying models to Kubernetes or SageMaker endpoints.
5 . Monitoring for data drift and model degradation.
6 . Handling distributed training (Horovod, Ray).
7 . Ensuring reproducibility (MLflow, DVC, GitHub Actions).
The Hardware Angle
Another important dimension is compute infrastructure:
1 . AI Applications: Often run on cloud APIs. You donโt worry much about GPUs, unless you want to self-host.
2 . ML/DL Models: Training deep networks can require powerful GPUs/TPUs (NVIDIA A100, H100, TPU v5).
3 . MLOps: Involves managing GPU clusters, storage, and networking (think: Kubernetes, Proxmox, or cloud GPU farms).
Final Thoughts
Think of it this way:
The ecosystem is huge and evolving fast. You donโt need to master everything at once. Pick the layer that excites you most, get hands-on, and then branch out as your curiosity grows.
At the end of the day, all three paths are connected โ many engineers eventually end up wearing multiple hats. Start where you are most comfortable, and build from there.
Continue reading...
