Deployment of AI Solutions

Hello students, welcome to Kwickprep. A model that works only on its maker's laptop helps nobody. How does an AI model reach millions of phones? Today we learn what deployment means, how a model becomes an app, why we monitor it, and real examples from India.

Let us first place deployment in the AI project cycle. We scope the problem and understand it. We acquire data. We explore the data to understand it. We build a model. We evaluate the model to check it is good enough. Only then do we reach the last stage, deployment.

Now, a new term. Deployment means putting a tested AI model into real use, where it actually helps people. After deployment, real people use it with real, everyday data, not just test data. And we deploy only after the model passes evaluation, because a weak model can harm people.

Here is what changes with deployment. Before, the model runs on a developer's laptop, and after, it runs inside an app or a website. Before, only the AI team uses it, and after, real users do. Before, it sees only test data, and after, it sees live, new data every day.

How does a model become an app people use? First, we save the trained model as a file. Next, we place it on a server, which is a powerful computer on the internet, or directly inside a phone app. Then we build an easy app around it, with buttons and screens. Before a full launch, we test it with a small group of users. Finally, we release it to everyone.

Let us follow one real kind of app, a crop disease app for farmers. A farmer photographs a sick leaf with a phone. The app sends the photo to the model. The model names the disease from the photo. The app then shows the disease and a remedy, in simple words, often in the farmer's own language.

A good model is not enough, the app must also be good. It must be easy to use, even for someone using a smartphone for the first time. It must be fast and give answers in seconds. It must be safe and protect users' personal data. And in India, it should work in the user's own language.

Is the work over after deployment? No, we must monitor the model, which means keep watching how it performs. Real data changes over time, for example, new kinds of spam messages appear. So the model's accuracy can slowly drop. The model may also face new situations it never saw in training. And users may report wrong answers that need fixing.

Monitoring works like a loop. We track accuracy and user feedback all the time. We ask, is the model still good enough? If yes, it keeps running. If no, we collect new, recent data. Then we retrain the model and evaluate it again. Finally, we deploy the updated model, and the watching starts again.

Pause and predict. A city's traffic prediction model was trained before a new metro line opened. After the metro opens, many people stop driving, and road traffic changes. What should the team do? They should monitor the drop in accuracy, collect new traffic data, and retrain the model.

Deployed AI is already all around you. Maps apps predict travel time from live traffic data. Email spam filters move spam messages away from your inbox. Banks use fraud alert systems that flag unusual payments, and may send you a warning message. DigiYatra, used at many Indian airports, lets passengers go through checks using face recognition. And Bhashini, a Government of India platform, uses AI to translate speech and text between Indian languages.

Finally, deployment must be responsible. Users should know when AI is making a decision about them. For important decisions, like health or loans, a human should check the AI's answer. And personal data must be protected, following laws like India's Digital Personal Data Protection Act.

Let us revise what we learned today. Deployment is the last stage of the AI project cycle. It puts a tested model into real use by real people. A model becomes an app through a model file, a server or phone, an app, a small test, and a release. After deployment, we monitor it, collect new data and retrain. And deployed AI is all around us, in maps, spam filters, fraud alerts, DigiYatra and Bhashini.