Chatbot Arena Human Preference Prediction

Chatbot Arena Human Preference Prediction

Kaggle Gold Medal

We trained Gemma-2-9B and Llama-3.1-8B with 4-bit QLoRA to predict human preferences, using AdamW and gradient accumulation to optimize log loss. To improve training efficiency and robustness, we tuned the context length and adopted length-sorted batching. We further improved performance by constructing a validation-weighted ensemble, yielding more stable and accurate preference probabilities.

Autonomous inventory management system

Autonomous Inventory Management System

Embedded Systems

Built an autonomous inventory management system that tracks items, expirations, and temperature in real time. Used an ATmega328p with a barcode scanner (UART), LCD, DS1631 sensor (I2C), and keypad, bridged to a Raspberry Pi over SPI. Integrated Firebase and a SwiftUI app for adding/removing items, viewing metadata, and receiving temperature alerts with instant two-way sync. Implemented a compact state machine for add/remove/info workflows with debounced inputs. Delivered hardware schematics, microcontroller firmware, a Pi client/server, and a complete mobile UI for autonomous, low-overhead inventory control.