A selection of ML/AI systems I've built — spanning space domain awareness, computer vision, generative models, and applied NLP.
Multi-model Resident Space Object (RSO) characterization pipeline for Space Domain Awareness. Integrates CNNs, UNet, and YOLO architectures to classify satellite components, estimate pose, and detect anomalies from monocular imagery in low-Earth orbit scenarios.
End-to-end semantic segmentation framework for high-resolution satellite imagery, targeting land-use classification and change detection. Uses a ViT-based encoder with a lightweight decoder head, trained on multi-spectral imagery and deployed via Kubernetes-backed inference endpoints.
Parameter-efficient fine-tuning (LoRA/QLoRA) of open-source LLMs on proprietary space-operations documentation and telemetry logs. Enables natural-language querying of mission data, anomaly summarization, and operator decision support with a retrieval-augmented generation (RAG) layer.
Novel-view synthesis pipeline for spacecraft Rendezvous and Proximity Operations (RPO) using 3D Gaussian Splatting. Generates photorealistic renders from sparse on-orbit imagery to augment training datasets for downstream pose-estimation models.
Physically Informed Neural Network (PINN) for high-resolution Synthetic Aperture Radar (SAR) image reconstruction. Incorporates physics-based loss terms derived from electromagnetic scattering models, achieving 88% prediction accuracy and reducing model hallucinations by 8%.