Hey everyone. I'm a CS student at FAST-NUCES in Karachi, and for the past 7 weeks I've been doing two things in parallel: the IBM Data Science Professional Certificate and an 8-week AI Engineering Internship that goes from Python fundamentals all the way to LLMs and deployment.
I wanted to write this partly to keep myself accountable, and partly because when I was starting out, reading other people's honest "here's what I actually struggled with" posts helped me a lot more than polished tutorials did.
So here's a real recap of the last 7 weeks — what I covered, what clicked, and what didn't.
Week 1 — Python for AI & Data Handling
Refreshed Python and got hands-on with NumPy, Pandas, and Matplotlib by building a Student Performance Analyzer on a class dataset. The interesting part wasn't the coding — it was seeing what the numbers actually said: pass rate came out to 93.3%, but the spread in Math scores was wide, clustering heavily at both the low (40s) and high (90s) ends rather than in the middle. Small reminder that "average" can hide a lot.
Week 2 — Machine Learning Fundamentals
Built a House Price Prediction model comparing Linear Regression, Decision Tree, and Random Forest on 200 records. Linear Regression actually won, with an R² of 0.9116 versus 0.86 and 0.85 for the tree-based models. That surprised me a bit going in — I expected Random Forest to dominate — but with only 200 rows, the more complex models didn't have enough data to show their usual advantage. Good early lesson that the "fancier" model isn't automatically the better one.
Week 3 — Deep Learning Basics
First real neural network, built in PyTorch for MNIST digit recognition. A simple 2-hidden-layer feedforward network hit 96.62% test accuracy after just 5 epochs, with training loss dropping from 0.38 to 0.09. Backpropagation finally clicked here — not the math itself, but watching the loss curve flatten out epoch by epoch made the "the model is learning" idea feel real instead of abstract.
Week 4 — Computer Vision
Moved to OpenCV and PyTorch for a Cat vs Dog classifier using transfer learning on a pretrained ResNet18 — only training the final layer instead of the whole network from scratch. This is where transfer learning really clicked for me: reusing a model already trained on 1.2 million images and just adapting the last layer felt like a shortcut that shouldn't work as well as it does.
Week 5 — Natural Language Processing
Built a Movie Review Sentiment Analyzer two ways: a ready-made Hugging Face pipeline (99%+ confidence out of the box), and my own fine-tuned DistilBERT on a 2,000-review IMDB subset, which reached 86.2% accuracy. The more useful result was actually a mistake — training loss kept dropping each epoch (0.33 → 0.05) while validation loss went up (0.42 → 0.58). Textbook overfitting, and the first time I actually saw it happen in my own numbers instead of just reading about it.
Week 6 — LLMs & Prompt Engineering
Explored how LLMs work under the hood and spent time experimenting with prompt design — small changes in phrasing or structure often changed the output more than I expected. Also got introduced to RAG (Retrieval-Augmented Generation) by building a Document Q&A Assistant, which combines a retrieval step (finding relevant chunks of a document) with generation (the LLM answering based on those chunks) instead of relying purely on what the model already "knows."
Week 7 — AI Engineering & Deployment (current)
Currently working with FastAPI and ChromaDB to build a Knowledge Base Chatbot API — essentially taking the RAG concept from last week and turning it into something that runs as an actual service instead of a notebook. This has been a different kind of learning curve: less about model accuracy, more about structuring an API properly, handling embeddings storage in a vector database, and thinking about how a real user (or another piece of software) would actually call this thing.
Wrapping up
That's the honest version of my last 7 weeks — some weeks went smoother than others, and the overfitting mistake in Week 5 probably taught me more than the weeks that went right. One more week to go before the capstone project, so I'll be posting an update once that's done.
If you're doing something similar — a certificate, an internship, or just learning ML/AI on your own — I'd genuinely like to hear how it's going for you. Feel free to drop a comment.
Top comments (2)
The most useful lesson here is that learning AI engineering is not just model theory. The day-to-day value comes from data cleanup, evaluation habits, small scripts that save time, and understanding where a model output can be trusted. That is a much more realistic picture than most internship recaps.
Really good point — I think I focused so much on the model-building side that I underestimated how much of the actual work was in areas like data cleanup and evaluation. Appreciate you pointing that out.