Five hands-on projects covering the AI/ML lifecycle end to end: data preprocessing, model development (transformers, CNNs, classical ML, LLMs/RAG), evaluation, and deployment. Each task lives in its own folder with a runnable pipeline, its own README (objective, methodology, results), and a Streamlit/CLI demo where applicable.
| # | Task | Core techniques | Key metrics |
|---|---|---|---|
| 1 | News Topic Classifier Using BERT | Transfer learning, transformer fine-tuning | Accuracy, weighted F1 |
| 2 | End-to-End ML Pipeline with Scikit-learn | Pipeline/ColumnTransformer, hyperparameter tuning |
Accuracy, precision, recall, F1, ROC-AUC |
| 3 | Multimodal Housing Price Prediction | CNN + tabular feature fusion | MAE, RMSE |
| 4 | Context-Aware Chatbot Using LangChain / RAG | Vector search (FAISS), conversational memory | Retrieval quality, source grounding |
| 5 | Auto-Tagging Support Tickets Using LLM | Zero-shot / few-shot prompting, fine-tuning | Top-1 / top-3 accuracy |
Fine-tunes bert-base-uncased on the AG News dataset to classify headlines
into World, Sports, Business, Sci/Tech. Tokenizes and preprocesses with
Hugging Face datasets/transformers, trains with the Trainer API,
evaluates with accuracy and weighted F1, and deploys as a live Streamlit or
Gradio demo.
Stack: BERT, Hugging Face Transformers, Streamlit/Gradio
Builds a single, reusable Scikit-learn Pipeline (preprocessing +
RandomForestClassifier) for customer churn prediction, so preprocessing
and modeling travel together as one object. Tunes hyperparameters with
GridSearchCV, evaluates with accuracy/precision/recall/F1/ROC-AUC, and
exports the fitted pipeline with joblib for direct inference on raw
records.
Stack: Scikit-learn, ColumnTransformer, GridSearchCV, joblib
Predicts housing prices by fusing a CNN image encoder (pretrained ResNet18) with a tabular MLP encoder (sqft, bedrooms, bathrooms, lot size, year built), concatenating both embeddings before a regression head. Evaluated with MAE and RMSE, plus a predicted-vs-actual scatter plot. Stack: PyTorch, torchvision (ResNet18), scikit-learn
A conversational chatbot that retrieves answers from a vectorized document
store (FAISS) and remembers conversation history via
ConversationBufferMemory, combined through a
ConversationalRetrievalChain. Deployed with Streamlit, showing which
source documents backed each answer. Supports a fully local LLM backend
(flan-t5-base) or OpenAI.
Stack: LangChain, FAISS, sentence-transformers, Streamlit
Tags free-text support tickets with their top-3 most likely categories, comparing zero-shot prompting, few-shot prompting, and a fine-tuned DistilBERT classifier head to head on the same dataset, with a side-by-side accuracy comparison chart. Stack: Transformers (DistilBERT), prompt engineering, scikit-learn
DevelopersHub-AIML-Internship-Advance-Task/
├── Task 1 News Topic Classifier Using BERT/
├── Task 2 End to End ML Pipeline with Scikit learn Pipeline API/
├── Task 3 Multimodal ML Housing Price Prediction Using Images Tabular Data/
├── Task 4 Context Aware Chatbot Using LangChain or RAG/
├── Task 5 Auto Tagging Support Tickets Using LLM/
└── README.md <- you are here
Each task folder is self-contained: its own requirements.txt, .gitignore,
and README.md with Objective / Methodology / Key Results sections, so any
task can be set up and run independently of the others.
Every task follows the same pattern:
cd "Task N <name>"
python -m venv .venv
.venv\Scripts\activate # Windows; use `source .venv/bin/activate` on Mac/Linux
pip install -r requirements.txtThen follow that task's own README for its specific data-generation, training, evaluation, and deployment commands.
- NLP & transfer learning (BERT, DistilBERT fine-tuning)
- Classical ML pipelines & hyperparameter tuning (Scikit-learn)
- Computer vision & multimodal fusion (CNNs + tabular data)
- Retrieval-Augmented Generation & conversational memory (LangChain, FAISS)
- Prompt engineering (zero-shot / few-shot) vs supervised fine-tuning
- Model evaluation across task types (classification, regression, ranking)
- Lightweight deployment (Streamlit, Gradio)
Muhammad Tayyab — GitHub