Tayyab Tanveer

Production system processing 2M+ rows from SSA (140 years) and Census datasets to compute 45+ linguistic attributes per name. Scraped metadata for 104K unique names including origin, meaning, and history. Calculates attributes using phonetic features (eSpeak) and LLM analysis with response caching. Optimized to run on 2GB RAM despite heavy phonetic libraries. Features similarity search with hybrid/visual (Jaro-Winkler)/phonetic (Double Metaphone)/semantic (sentence transformers + FAISS) algorithms with origin filtering. Async FastAPI handles 150+ concurrent users (Locust-tested). Migrated from CSV to Parquet for faster searches and lower memory overhead. Includes API key authentication and Docker deployment.

LangGraph-powered agent that converts natural language queries into SQL over an enterprise retail DB with 350+ tables across multiple branches. Built MVP with 30 tables using M-Schema method to reduce hallucinations on table values/columns, adding contextual descriptions for better table selection. Evolved from LangChain to LangGraph agents, switching from Llama Maverick to GPT-OSS 20B, reducing latency from ~10s to 3-5s. Includes FastAPI endpoints for text and voice queries with Whisper 3 Turbo (STT) and Google/OpenAI TTS.

Automated system that generates podcast scripts via LLM, converts to speech with Play.ht, and produces dual-avatar video podcasts using Pyannote diarization and SadTalker. This is a test project to explore the potential in AI-generated video podcasts, not an MVP or production-ready system.

Session-based PDF chatbot with stateful conversation history. Processes PDFs into vector embeddings using Chroma, enabling contextual Q&A powered by Google's Generative AI with persistent chat sessions per user.

Real-time conversational avatar system for small businesses. Simplified the full STT→LLM→TTS→Avatar pipeline end-to-end, cutting response latency from 3–5 minutes down to 6-8 seconds.

K-Means clustering analysis on customer purchasing behavior. Identified 3 distinct segments using Elbow Method optimization, revealing patterns in purchase frequency and monetary value for targeted marketing strategies.

Multi-label toxicity classification system using Bidirectional LSTM. Detects 6 toxicity categories (toxic, severe toxic, obscene, threat, insult, identity hate) with web-based prediction interface via Anvil integration.

ML-based network IDS trained on KDD99 and NSL-KDD datasets. Achieved 100% accuracy on KDD99 using Random Forest with SMOTE balancing and Chi-squared feature selection, comparing performance across 5, 10, and 15 feature subsets.