01 / THE LAB
SELECTED
WORK
Experiments where models, interfaces and human questions collide. Some polished. Some still lab notes. All built to learn something real.
DHAKA / AVAILABLE WORLDWIDE
“I can't stop thinking about tomorrow.”
A multi-label emotion detector using the GoEmotions taxonomy and a DistilBERT-based approach, delivered as an interactive Streamlit experience.
After finishing a machine-learning course, I wanted to go deeper than the coursework. I chose GoEmotions—27 emotions plus neutral—as the foundation and trained a DistilBERT model to classify emotion from text, saving the model and tokenizer with Hugging Face’s save_pretrained.
The backend, from data preparation to training and model persistence, was built from scratch. AI assistance accelerated the interface. Early predictions appeared as raw labels such as class_9; mapping those labels back to real emotion names let the app finally speak human: “You are feeling joy.”
CASE FILE 02 / STRUCTURED DATA
CREDIT
SCORE ML
ACCURACY / %
A large structured-data prediction project built around feature engineering, model comparison and ensemble learning.
The work tested multiple approaches against the same problem, then combined their useful behavior through stacking. The best stacking result reached approximately 0.798 accuracy, with a Streamlit front end turning the experiment into an interface people can actually use.
Beyond a single score, the case demonstrates the full working loop: explore the data, transform the signal, train competing models, evaluate honestly, and give the result a usable surface.
CASE FILE 03 / RECOMMENDATION
MOVIE
RECOMMENDER
Quiet tension.
Strange worlds.
Earned endings.
A hybrid recommendation system combining content similarity, collaborative signals and learned methods.
TF-IDF interprets what a film is about. User-item kNN looks for patterns in taste. Random Forest methods add another predictive layer. Together they form a recommendation system that is more interesting than any one approach on its own.
The project reflects the way I prefer to build: use multiple perspectives, compare their blind spots, then design the handoff between them.
MORE EXPERIMENTS / STILL IN MOTION
Smaller builds, same curiosity.
THE LAB
NEVER CLOSES.
HOW I WORK WITH AI / NO MAGIC WAND
FASTER
EXPERIMENTS.
HUMAN
JUDGMENT.
I use generative AI as a creative and development accelerator: exploring concepts, prototyping interfaces, iterating visual directions, debugging and supporting implementation. The goal is not “AI-generated for the sake of AI”; it is faster experimentation paired with technical judgment and a deliberate final experience.