Projects

Selected work spanning ML infrastructure, computer vision, and multimodal AI.

Production Level Inference Server

LLM-focused inference server in FastAPI with ONNX runtime graph optimizations. Deployed on AWS EC2 with K8s via Terraform.

FastAPIONNXTerraformK8s

Short Story Generation from a Single Image

Image synthesis system using CLIP embeddings from reference images, extrapolating them to create a visual narrative with chronologically and logically related images.

CLIPDiffusionPython

Hateful Meme Detection

Facebook's Hateful Memes competition — classified memes as hateful or non-hateful using multi-modal techniques and data augmentation on fine-tuned LLMs. Achieved rank 1 at peak, top 50 finish.

Multi-modalLLMsPyTorch

Speed Tracking (YOLO-v3)

Real-time speed estimation using YOLO-v3 and a centroid tracker. Predictions at 30 fps.

PyTorchYOLOComputer Vision

DeepFake Detection

Kaggle Deep Fake Detection challenge — frame extraction with ffmpeg, face detection with RetinaFace, classification with EfficientNet.

EfficientNetRetinaFaceKaggle

Logo Detection

Logo detection in video clips using EfficientDet on the Openlogo dataset. Achieved 70.5% IoU and 85% label accuracy.

EfficientDetPyTorchObject Detection