Most AI projects never make it to production. The gap between a working notebook and a reliable product is enormous - spanning infrastructure, monitoring, team processes, and cost management. This lesson covers everything you need to bridge that gap successfully.
Before writing a single line of code, ask the most important question in AI product development: does this problem actually need AI? The answer is "no" more often than you would think.
AI is the right choice when:
AI is the wrong choice when:
The 10x rule: If AI does not deliver at least a 10x improvement over the non-AI alternative in speed, accuracy, or cost, the operational complexity of maintaining an ML system is rarely worth it.
Not every problem needs a large language model. Choosing the wrong model type is one of the most expensive mistakes in AI product development. Match the tool to the task:
| Problem Type | Best Approach | Example | |---|---|---| | Tabular prediction | Gradient boosted trees (XGBoost, LightGBM) | Churn prediction, pricing | | Image tasks | CNNs or vision transformers | Quality inspection, OCR | | Text classification | Fine-tuned BERT or smaller LLMs | Sentiment analysis, routing | | Generative text | LLMs (GPT-4, Claude, Llama) | Chatbots, content generation | | Time series | Prophet, N-BEATS, or temporal CNNs | Demand forecasting |
For a tabular churn prediction task with 50,000 rows of structured data, which approach is typically most effective?
Sign in to join the discussion
Speed of iteration beats perfection every time. Teams that spend months building before validating waste resources on assumptions. Use this framework to validate ideas quickly:
Day 1 - Scope and data. Define the single metric that matters. This is non-negotiable - without a clear success metric, you cannot evaluate anything. Gather or simulate the minimum viable dataset. Do not clean it perfectly - just enough to be usable.
Day 2 - Build and test. Use an existing model or API. Wrap it in a minimal interface (Streamlit, Gradio, or a simple API). Get something a human can interact with.
Day 3 - Validate. Put it in front of real users or stakeholders - not colleagues who will be polite, but people who would actually pay for the solution. Measure against your success metric. Decide: kill, pivot, or invest.
Think about the last AI project you worked on. How long did it take to get the first version in front of a user? Could a 3-day sprint have validated the core idea faster?
Production ML requires rigorous operational practices that most data scientists never learn in academic settings. The core principle is deceptively simple: version everything, automate everything. The execution, however, requires discipline and the right tooling.
What to version:
Pipeline automation:
Google's research found that only about 5% of real-world ML system code is the actual model. The remaining 95% is data pipelines, serving infrastructure, monitoring, and configuration management.
Deploying a model is not the finish line - it is the starting line. Without proper monitoring, models degrade silently, and by the time someone notices, the damage is already done.
Data drift - The input data distribution shifts over time. A model trained on pre-pandemic shopping data will fail on post-pandemic patterns. Monitor feature distributions continuously and alert on statistically significant shifts using tests like the Kolmogorov-Smirnov test or Population Stability Index.
Model degradation - Even without data drift, model performance can decay as the real world changes. User behaviour evolves, competitors shift the market, and seasonal patterns alter outcomes. Track prediction confidence, error rates, and business metrics continuously.
Key monitoring practices:
What is 'data drift' in the context of production ML systems?
AI inference costs can spiral quickly. Smart architecture decisions keep them manageable.
Practical strategies:
Which cost optimisation strategy involves using different models based on query complexity?
If you were given a budget of £10,000/month for AI inference, how would you allocate it? Consider which parts of your system need real-time responses versus batch processing, and where caching could eliminate redundant computation.
Building AI products is fundamentally a product and engineering challenge, not just a machine learning one. Keep these principles close:
The best AI products are built by teams that understand both the technology and the business problem deeply. Technical excellence without product thinking leads to impressive demos that no one uses.