page 02 — sketches that became code

Projects

SPACE-HOP

Spacecraft 6-DoF Pose Estimation using Embedding-Predictive Pretraining and Hopf Map

First application of Joint-Embedding Predictive Architectures (JEPA) to the spacecraft domain.
Keypoint masking strategy that forces the network to learn deep 3D structure under severe orbital occlusions.

Tournament Inpainting for Patch-level Alignment in text-to-image diffusion models with Tournament Sampling Policy Optimization

TiPAI: a decoding-time alignment framework for diffusion models using localized timestep-aware auditing and targeted inpainting for policy-compliant generation.
TSPO: (Tournament Sampling Policy Optimization), a lightweight reinforcement learning framework for selecting optimal inpainting configurations under quality and compute constraints.

Dense point tracking + LiDAR perception

Multi-modal 3D perception stack integrating LiDAR and depth cameras, executing pointcloud alignment, auto-calibration using SuperGlue, and human segmentation (PointNet++) within a ROS/RViz2 environment.
Macroscopic 3D tracking with a dense optical flow pipeline (AllTracker/RAFT) and a custom statistical metric (Directness Ratio) to isolate flow vector variance, successfully distinguishing micro-scale involuntary tremors from voluntary locomotion.

Self-Hardening Transformer Web Application Firewall via Incremental Learning

Built a firewall that catches zero-day attacks utilizing a DistilBERT encoder trained via Masked Language Modeling (MLM) for reconstruction loss, coupled with a frozen contrastive [CLS] head to detect zero-day anomalies with 93.78% recall.
Red-Team self-hardening loop was designed where a fine-tuned Llama agent interprets high-loss anomalies to autonomously generate and inject regex signatures into the static engine, instantly patching new exploits.