Ankur Praveer

Market Basket Analysis Using R, VIT University, Vellore, Tamil Nadu  05/01/21

The 'Market Basket Analysis using R' project focused on applying the Apriori and Eclat algorithms to identify frequent itemsets in retail transaction data. These itemsets are groups of products commonly purchased together. By generating association rules from these itemsets, we identified strong relationships between different products. The analysis provided valuable insights for business applications, particularly in developing effective cross-selling strategies.

PoMa: Efficient Multi-Token Prediction through Positional Matrix - Paper - April 2025
Supercut: AI Video Segment Retrieval and Summarization System
Built an AI-powered system that generates video “supercuts” by automatically identifying, extracting, and stitching semantically related
clips from YouTube or meeting recordings from user based query.
Hybrid Movie Recommendation LLM Based Agent
Implemented a Hybrid Learning-Fuzzy Logic Model for Time-Series Prediction, Nanyang Technological University, Singapore
Search Engine & Review Summarization System, Nanyang Technological University, Singapore
Built an advanced search engine and review summarization tool using PyLucene, NLTK, and Docker.
Developed a lightweight framework for multi-token generation using learnable positional adapters on frozen LLaMA decoders,
achieving up to 3.2× inference speedup and 67.97% top-5 accuracy on CodeAlpaca and CoNaLa without full model fine-tuning.
•
Implemented k-positional parallel decoding, predicting multiple tokens in a single pass to significantly enhance LLM inference
speed while maintaining scalability and architectural integrity.
•
Designed an end-to-end FastAPI pipeline integrating LangChain, Gemini, WhisperX, ASR, and FAISS for multimodal retrieval using
transcripts and OCR-extracted text.
•
Integrated RAG (Retrieval-Augmented Generation) for contextual summarization and scene clustering using Gemini embeddings
and LLM-based captioning.
•
• Optimized concurrent processing with asynchronous Python tasks, Redis caching, and modular APIs for scalable deploymen
Designed and implemented a scalable, modular recommendation engine combining collaborative filtering and content-based
filtering in Python, achieving 15% higher accuracy over baseline CF.
•
Integrated a retrieval-augmented generation (RAG) pipeline with live web search and LLM-powered summarization
(DeepSeek/OpenAI) to enrich movie metadata beyond static datasets using LangChain, HuggingFace and FastAPI.
•
Developed a hybrid Fuzzy LSTM model achieving up to 3× improvement in portfolio return stability over baseline LSTM on 15 years
of REIT stock data.
•
• Applied transfer learning with selective layer fine-tuning, improving adaptability to new market regimes with minimal retraining.
• Implemented fuzzy, and geospatial search with BM25 ranking for typo-tolerant and location-based queries.
• Developed sentiment analysis and key phrase extraction to summarize reviews by frequency and user sentiment.