Vision Intelligence Infrastructure: A high-performance, async control panel for human-in-the-loop AI annotation.

A sleek, dark-themed control panel designed for decoupled ML microservices and robust task queues, eliminating UX bottlenecks with pure speed and instantaneous rendering.

Key Innovations:

• 🚀 Asynchronous ML: Non-blocking AI auto-labeling via PyTorch, LayerCAM & YOLO.
• ⚡ Zero-Latency UI: Hardware-accelerated React-Konva staging canvas.
• 🔄 Decoupled Workers: Infinite horizontal scaling using Celery message brokers.
• 🔒 Isolated Workspaces: Robust Role-Based Access Control (RBAC) circuits.

"Finding the needle in the haystack, from 400ft above."

A novel coarse-to-fine computer vision pipeline designed for efficient small object detection in high-resolution (2K/4K) aerial imagery. Tackles the critical trade-off between resolution and latency in drone forensics.

Key Innovations:

• Uses LayerCAM to identify semantic "hotspots" before processing.
• Intelligently slices and zooms into regions of interest—skipping 80%+ of empty backgrounds.
• Outperforms blind sliding-window approaches (SAHI) in both speed and accuracy.

Transform any image into a masterpiece — in real-time.

A fast neural style transfer implementation that generates stylized images using a feed-forward CNN trained with perceptual loss. Performs instant stylization in a single forward pass.

Key Features:

• 🚀 Real-time inference with a custom residual architecture.
• 🧠 Perceptual content & style loss using a pretrained VGG-16 network.
• 🔁 Instance Normalization integrated for high-quality, artifact-free outputs.
• 📦 ONNX export supported, ready for edge deployment.

🧭 pygog (Google CLI Agent)
A powerful CLI for Google services (Gmail, Drive, Calendar). Features a built-in natural language AI agent supporting Gemini, DeepSeek, & OpenAI.
<Python> <Google APIs> <LLM Agents>

📐 Depth Estimation + Semantic Seg.
Multi-modal depth completion using RGB + sparse depth + semantic maps. Features a DepthNet-style encoder-decoder trained on NYU Depth v2 with multi-scale supervision.
<PyTorch> <NYU-Depth-v2> <Encoder-Decoder>