Farjana Sultana Samia

Featured Projects

• RAG Assistant

  GitHub Demo video
  • Built a five-stage retrieval-augmented generation pipeline: document ingestion, chunking at 400 characters with 50-character overlap, local sentence-transformer embeddings, top-4 vector retrieval with FAISS, and answer generation through the Claude API.
  • Wrapped the pipeline in a Funix web application that accepts .txt, .csv and .pdf uploads, runs batch evaluation against sample question–answer pairs, and shows the retrieved source chunks alongside every answer.
  • Added token-usage and cost tracking and packaged the system with Docker. Embedding and retrieval run locally, so only generation calls consume API credits.

  

  

  Tech

  • Python
  • Claude API
  • FAISS
  • sentence-transformers
  • LangChain
  • Funix
  • Docker
  • RAG

• Token-efficient RAG with query-aware summarization

  Slides
  • Standard RAG sends whole retrieved chunks to the LLM, so most prompt tokens have nothing to do with the question. This project cuts that cost for high-volume HPC question-answering workloads (CS 6250).
  • Added a two-stage compression step after retrieval and re-ranking: keyword-overlap filtering keeps the top-k sentences of each document, then a local T5 model writes a query-aware summary that replaces the full retrieved text in the GPT-3.5-Turbo prompt.
  • On 918 questions from rag-mini-wikipedia, T5-Small summarization used 60% fewer tokens than the baseline while exact match fell only 2.9 points and BERTScore stayed above 0.93.
  • 60%fewer tokens with T5-Small (604K → 241K)
  • −2.9points exact match (58.28% → 55.34%)
  • 95%of baseline accuracy kept
  • >0.93BERTScore for every variant
  Method
  1Retrieve 2Filter 3Summarize 4Prompt 5Generate

  The same pipeline as the baseline: documents are cleaned, chunked and embedded, then FAISS dense search finds candidate chunks and a cross-encoder re-ranks them.

  • PDF ingestion (PyPDF2)
  • Text cleaning
  • 512-token chunks
  • BAAI/bge-large-en-v1.5 embeddings
  • FAISS search
  • Cross-encoder re-ranking
  • Top 5 documents

  From each retrieved document, keep the top 5 sentences by word overlap with the query, scored as |s ∩ q| / |s ∪ q|. If no sentence overlaps, fall back to the first 5 sentences.

  A local FLAN-T5 model condenses the filtered sentences into a 120–180-token query-aware summary, prompted with “Summarize the following text in context of the question”. Three sizes were compared:

  • flan-t5-small (60M)
  • flan-t5-base (220M)
  • flan-t5-large (770M)

  The final prompt is query + summaries instead of query + full retrieved text, which is where the token savings come from.

  GPT-3.5-Turbo answers from the compressed prompt. Summarization runs locally, so it adds no API cost, and the design works with any generator model.

  

  

  Results on 918 rag-mini-wikipedia questions

  Exact match F1 BERTScore Tokens

  Bigger summarizers did not help: the 770M-parameter T5-Large scored lowest and saved no more tokens than the 60M-parameter T5-Small, likely because it over-compresses and drops key facts. Average prompt size fell from 658 to 262 tokens per question.

  Performance by question type

  Baseline T5-Small T5-Base T5-Large

  Exact match F1

  Yes/No questions, 46% of the dataset, stay above 83% for every model. Why questions stay at 4% or below, and What questions, a quarter of the dataset, sit near 27%: the clearest targets for future work.

  Tech

  • Python
  • FAISS
  • Cross-encoder re-ranking
  • FLAN-T5
  • LangChain
  • PyTorch
  • Hugging Face Transformers
  • Nova HPC (2× A100)
  • GPT-3.5-Turbo
  • BERTScore
  • ROUGE
  • RAG

• Aggregated Token Attention Network (ATAN) for food review sentiment

  Slides
  • Three-class sentiment classification (negative, neutral, positive) of Amazon Fine Food Reviews, combining review text with user-specific rating data.
  • Fine-tuned several pre-trained BERT models and averaged their token-level attention scores into a single aggregated attention, which drives the final class prediction.
  • The aggregated model reached 77% accuracy and 0.77 macro F1 on a balanced 12,000-review test set, ahead of every individual model.
  Method
  1Preprocess 2Balance 3Fine-tune 4Aggregate 5Predict

  Each review is cleaned before tokenization:

  • Punctuation removal
  • URL and unwanted-word removal
  • Lowercasing
  • Stop-word removal
  • Lemmatization

  The raw data is heavily skewed toward positive reviews, so each class is down-sampled to 20,000 reviews, giving 60,000 in total, then split 80:20 into 48,000 training and 12,000 test reviews.

  Pre-trained BERT models are fine-tuned on the review text and user ratings:

  • bert-base-uncased
  • lxyuan/distilbert-base-multilingual-cased-sentiments-student
  • nlptown/bert-base-multilingual-uncased-sentiment

  Token-level attention scores from every model are averaged into one aggregated attention. Each model ships its own tokenizer, so the attention vectors differ in length and are padded to a common size before the element-wise average.

  The aggregated attention feeds the classifier, which outputs one of three labels: Negative (0), Neutral (1) or Positive (2).

  

  Results on the 12,000-review test set

  Accuracy Precision Recall F1

  Macro avg Negative Neutral Positive

  Tech

  • Python
  • PyTorch
  • Hugging Face Transformers
  • BERT
  • DistilBERT
  • Attention
  • Sentiment analysis

Machine Learning & Computer Vision

• Red blood cell counting from blood-cell images

  GitHub
  • Object-detection models that locate and count red blood cells in microscopy images.
  • Compared Faster R-CNN, SSD and YOLO using transfer learning from pre-trained backbones.
  Tech

  • Faster R-CNN
  • SSD
  • YOLO
  • Transfer learning
  • Object detection

• Flower recognition with residual networks

  GitHub
  • Image classifier over 4,242 flower photographs in five classes — chamomile, tulip, rose, sunflower and dandelion — with roughly 800 images per class and varying resolutions.
  • Fine-tuned ResNet50 on the dataset, reaching 90% classification accuracy.
  Tech

  • ResNet50
  • Keras
  • Transfer learning
  • Image classification

• CIFAR-10 object recognition

  GitHub
  • Convolutional image classifier over the CIFAR-10 dataset: 60,000 colour images at 32×32 resolution across 10 object classes.
  • Implemented and trained in Keras.
  Tech

  • Keras
  • CNN
  • Image classification

Data Engineering

• Web scraping and data collection

  Twitter Amazon Bengali news
  • Twitter profile scraper that extracts public profile information, built with Selenium.
  • Location-based Amazon product scraper that collects product information across regions, built with Selenium.
  • Bengali news scraper covering prothomalo.com, banglatribune.com, jagonews24.com and samakal.com, built with Scrapy.
  Tech

  • Python
  • Selenium
  • Scrapy
  • Web scraping

Web Development

• Semester grading system

  GitHub
  • Web application for recording and calculating semester grades, built with Python Flask.
  • Interface designed with Bootstrap.
  Tech

  • Python
  • Flask
  • Bootstrap