Gayathri Adulla

A multimodal deep learning system fusing DistilBERT text embeddings with ResNet-18 visual features to predict movie genres from posters and plot overviews, deployed via an interactive web interface.

• Dual-branch late-fusion architecture aligning text and image latent features
• Addressed class imbalance with weighted BCE loss and label co-occurrence analysis
• Evaluated with macro/micro F1, confusion matrices, and threshold tuning

A hybrid deepfake detection framework combining unsupervised representation learning with supervised latent-space classification — detects synthetic faces by modeling structured latent distributions rather than surface-level pixel artifacts.

• β-VAE (β=4) with 128-dimensional latent space, trained on FFHQ real faces vs. Stable-Diffusion-generated fakes
• KL-divergence anomaly detection combined with logistic regression on standardized embeddings
• 98% classification accuracy, 0.997 ROC-AUC, t-SNE-validated class separability

A containerized, cloud-native runtime orchestrating autonomous AI agents — a centralized FastAPI orchestrator coordinates task routing and execution across independent agent microservices via a Redis message queue.

• Simulates 50+ concurrent agents with health checks, retry logic, and structured JSON logging
• Modular architecture for rapid integration of new agent types
• Demonstrates cloud-native patterns used in modern agentic AI platforms

A research pipeline reconstructing full 3D human body meshes from monocular video using SMPL-X parametric models, with a CPU-optimized rendering pipeline and a Gradio-based comparison interface.

• PyTorch tensor pipeline integrated with Trimesh mesh manipulation and Pyrender physically-based rendering
• Converts raw NPZ pose/shape parameters into frame-wise rendered, vertex-normalized mesh outputs
• Vectorized batch processing to minimize memory overhead without GPU dependency

A classical computer vision pipeline recognizing faces under heavy occlusion — eyes-and-nose-only inputs simulating masks, low-quality footage, or cropped images.

• Fused HOG, SIFT, LBP, and color histogram features with PCA dimensionality reduction
• SVM classification combined with cosine/Euclidean similarity matching
• Ensemble feature-fusion strategies improved robustness under rotation, scale, and lighting variation

A Java desktop application managing multi-role e-commerce workflows — Admin, Customer Support, Warehouse, Delivery, Supplier, and Fraud Analyst — with role-specific dashboards and a normalized MySQL backend.