This AI fact-checking system, built with LangGraph, dissects text into verifiable claims, cross-referencing them with real-world evidence via web searches. It then generates detailed accuracy reports, ideal for combating misinformation in LLM outputs, news, or any text.

From-scratch, paper-faithful PyTorch implementation of DeepSeek-V4 architecture for transparent study, testing, ablation, and mini-scale training.

chromaVive turns grayscale images into vibrant, colorized versions using advanced deep learning inspired by Richard Zhang's research at UC Berkeley. Powered by models from the ECCV16 and SIGGRAPH17 papers, chromaVive advances the art of image colorization.

This is a Research Paper implementation of Levi-Hessaner Age-Gender Classification model which has re-implemented and trained on Indian faces.

Data structures & algorithms project for implementing the Hilbert R Tree data structure as present in the original research paper, focusing on the adding of elements and searching of elements along with a pre-order traversal of the tree.

End-to-End Python implementation of CompactPrompt (Choi et al., 2025): a unified pipeline for LLM prompt and data compression. Features modular compression pipeline with dependency-driven phrase pruning, reversible n-gram encoding, K-means quantization, and embedding-based exemplar selection. Achieves 2-4x token reduction while preserving accuracy.

From-scratch PyTorch implementation of the Gated Attention mechanism for transformer architectures, focusing on efficient sequence modeling and attention optimization.

QueensGambit is a Python-based chess engine using the Minimax algorithm with alpha-beta pruning to make optimal moves. It incorporates piece values inspired by AlphaZero and an entropy-based evaluation for dynamic gameplay. The engine supports both human-AI and AI-AI matches.

End-to-End Python implementation of Semantic Divergence Metrics (SDM) for LLM hallucination detection. Uses ensemble paraphrasing, joint embedding clustering, and information-theoretic measures (JSD, KL divergence, Wasserstein distance) to quantify prompt-response semantic consistency. Based on Halperin (2025).

This project is my PyTorch reproduction of PaliGemma, a compact 3B vision–language model that integrates SigLIP vision features with a Gemma decoder. I implemented the full multimodal pipeline from vision encoding to autoregressive text generation to study modern VLM architectures from a research perspective.

Production implementation of OPRO (Wei et al., 2023) — gradient-free prompt optimization via LLM-as-optimizer. Implements p* = argmax f(p) over discrete natural language space. Built from the paper, no prompt libraries. arXiv:2309.03409

Build enterprise multi-agent customer service systems for finance and e-commerce with Python, Java, and Go implementations including interview preparation materials.

📈 Analyze how rumors affect investment decisions using a qualitative reasoning engine for better outcomes and informed choices.

This project is a clean, from-scratch implementation of a GPT-2 style autoregressive transformer built using PyTorch. Unlike character-level toy models, this implementation operates on tokenized text (via tiktoken) and supports modern training features.

Research-grade energy forecasting pipeline for smart homes: rigorous data cleaning (IQR outlier detection, temporal imputation), normalized MySQL database (3NF), scientifically validated feature engineering, and CPU-optimized ML models (XGBoost R²=0.938, MAPE<10%)

End-to-End Python implementation of a bankruptcy prediction method which adapts Altman's Z-Score to Compositional Data Analysis (Keivani et al., 2026). Benefits: Uses the Aitchison simplex to eliminate outliers & asymmetry in financial ratios. Pipeline: log-ratio EM imputation, pairwise log-ratios, ML classifiers (Logit/k-NN/RF), validation.

markoText is a Python-powered story generator using a Bag of Words Markov Chain model to craft narratives. By training on input text, it learns word transition probabilities to generate unique content mimicking the style and structure of the original.

evoGrad is a lightweight automatic differentiation engine designed to help users understand autograd concepts and neural network operations. It offers a minimal framework for gradient computation and backpropagation, making it an ideal tool for beginners to explore deep learning fundamentals.

This project is a minimal implementation of a character-level language model inspired by the makemore series by Andrej Karpathy. The goal is to build an intuitive understanding of how neural networks can learn patterns in text and generate new sequences—in this case, human-like names.

SpeechCueLLM Implementation: Enabling Llama-3 to detect emotions from speech vocal nuances via text description mapping. | Python, PyTorch, HuggingFace, Llama-3, LoRA, Praat | AI, NLP, Multimodal Speech Emotion Recognition