Ayan Chowdhary

Built a crop intelligence system for a Maryland-based CSA to reduce overplanting and improve profitability. Transformed planting and field data into insights on crop performance, then integrated sales transaction data in Databricks to analyze demand versus supply. Enabled data-driven decisions on what to grow, where to plant, and how much.

Built an autonomous soil-analysis rover that integrated pH, moisture, EC, and temperature sensors with GPS and gyroscope correction. Streamed real-time field data over MQTT to a Mapbox dashboard and collaborated with local farms and University of Maryland CPSE to support practical agricultural decision-making.

Developed a web app to increase recycling awareness in Montgomery County through educational resources and interactive experiences. Built the project with Next.js, Tailwind CSS, and TypeScript to create a responsive and accessible experience.

Founded a computer science club at school to help students build projects, prepare for competitions, and grow as technical collaborators. Organized the group around leadership, public speaking, and hands-on programming, with plans to compete at nearby universities including Penn and UMD.

Built an AI-powered app that generates a custom Pokemon card from a user prompt in under 20 seconds. Integrated Firebase for backend services and OpenAI for generation workflows, then deployed the product using a modern full-stack web stack.

Placed in the top 71st percentile on Kaggle in a competition focused on predicting smoker status from bio-signals. Trained and tuned XGBoost and LightGBM models with Optuna to improve performance through structured experimentation.

Built a taxi simulator using reinforcement learning to explore state, action, and reward dynamics in practice. Trained a Q-learning agent to optimize passenger pickup and drop-off behavior while deepening my understanding of decision-making systems.

Built an early GPT-powered application that combined language models with Google web scraping to surface timely answers. Added content filtering, blocklist management, and custom demos as part of a beta testing workflow.

Built a car-image generation project inspired by DC-GAN techniques and paired it with a super-resolution model to enhance low-resolution outputs. Trained models with adversarial loss on car datasets while exploring deeper image-generation workflows.

Built a voice interface for GPT-3 that supported spoken interaction and audio responses. Implemented the project in Python and explored speech recognition and speech synthesis as part of an early conversational AI workflow.

Implemented an advanced CNN image-classification model based on the ResNet architecture paper using TensorFlow and Keras. Focused on large-scale image tasks and model optimization while learning modern deep learning practices.

Studied transformer architectures such as GPT-2 and BERT, comparing their performance with earlier natural language processing approaches. Explored tokenization, embeddings, and model training to better understand modern language systems.