Built an AI workflow from scratch, integrated it with iOS UI, and captured 28% of students in a high-school pilot, highlighting strong appeal and engagement.
Built personalized AIs with unique personalities that interact, connecting users and fostering deeper engagement. Led the entire project from start to finish.
Boosted sales efficiency by 30% using GPT-3.5, RAG, and prompt chaining to automate responses and project estimates, tracking via LangChain & LangSmith.
Developed AR-based room size estimation, securing an extra $2M in funding.
Developed an object detection system using Grounding DINO and Segment Anything, driving $4M funding and a Y Combinator launch.
Built a transformer-based architecture to convert time series into images using autoencoders and tested pix2pix for spectrogram conversion.
Conducted knowledge distillation, combining model-generated and human-labeled data to improve BRT's production model performance significantly.
Developed a high-accuracy Segformer model for crop/weed segmentation, outperforming the current model. BRT may adopt it after hardware upgrades.
We propose a deep Q-learning chess AI using rCNN and qCNN to evaluate board states and compare its efficiency to classical CNNs in real-time move selection.
Developed a model predicting COVID vaccination status for targeted marketing, ranking 16th out of ~700 teams in a Humana data analytics project.
We used camera parameters to re-identify people across multiple camera views, as deep learning couldn't be applied due to identical clothing.
A minimal AlacGan implementation based on the User-Guided Deep Anime Line Art Colorization paper and GitHub repository.
I generated scenes with people in various poses by using a segmentation CNN, placing them logically, and adapting HSV with histogram equalization.
I built a full AI pipeline, using website APIs and StyleGAN2, and started selling AI-generated dragons as NFTs to learn about crypto.
I automated loader bucket teeth counting using template matching and color histograms, ensuring none were missing from the video samples.
I researched contrastive learning methods like SimCLR and Swav, using ResNet, ResNext, and EfficientNet for image retrieval and object detection.
I built a CNN-based multilabel classifier that detects language and speaker sex from audio mel-spectrograms.
The project automates dataset creation with OpenCV webcam tracking and fine-tuning, followed by an object detector using the fine-tuned model.
In this competition, we had to predict satellite position and velocity. My contribution was an attempt to solve the task by using gradient boosting and system