Shiv Gupta

Automated cross-platform arbitrage bot exploiting pricing inefficiencies between Stake (sportsbook) and Kalshi (prediction market) for IPL cricket. Devigging, 12-step execution protocol, and risk-managed shadow/live modes.

Interactive tool for random person picking, team generation, and one-to-many people organization with smart diversity-maximizing shuffles, lock-in preferences, and password-protected saveable boards.

Competitive AI poker agent for the Jump Trading x CMU AI Poker Tournament 2026. Features exact equity computation, Bayesian opponent modeling, Monte Carlo sampling, and position-aware strategies over a modified 27-card deck.

Training-free latent reasoning for frozen LLMs. +13pp on GSM8K via KV-cache recurrence, split-layer generation, and answer-mass gating — zero training, zero fine-tuning.

Replication and extension of the 'Who is Harry Potter?' unlearning experiment - selectively erasing specific knowledge from LLMs while preserving general capabilities.

Privacy-first desktop app with an LLM orchestration layer using Ollama, engineering structured prompt chains and output parsing to generate meeting notes from Whisper-transcribed audio - zero external API dependencies.

Developed Diffusion and ControlNet based Interior & Exterior design generation models at the National University of Singapore. Research on applying generative AI techniques for architectural visualization and design under the guidance of Dr Tan Wee Kek.

End-to-end autonomous navigation system using ROS with PID controllers, computer vision pipelines, and path planning for real-time beacon identification and obstacle avoidance.

Built a graph crawler mapping YouTube's recommendation algorithm across 10K+ videos with ForceAtlas2 community detection. Published as 'Parametric Algorithmic Transformer Based Weighted YouTube Video Analysis' - a custom weighted scoring framework with modularity-based classification.

Architectured and designed code for models like CNNs, and developed models using skeleton codes of ResNet152 and VGG19 for detecting fire and smoke in visual input. Collected data, preprocessed it, augmented it to prevent skewness and used this custom dataset for training. Hyperparameter tuned the model based on results and deployed it using Streamlit.

A macOS-themed terminal personal website with draggable windows, embedded Spotify and YouTube players, and an interactive command-line interface.

Developed the Frontend and Backend for an app where users can find the nearest EV swap/charging station and book a swap while on route. Developed AI analysis for optimal map positions of EV Swap Stations and a mathematical algorithm for pricing to solve battery swap mismatch through an AI model trained for battery degradation prediction.