ARSH ABBAS
NAQVI
AI researcher at Delhi Technological University, working on world models for embodied AI, safer generative systems, and AI safety research.
◇ ABOUT
EDUCATION: B.Tech, Computer Science & Engineering — Delhi Technological University.
AFFILIATIONS: UAS-DTU · ML Research Lab, DTU · Pragya Lab, BITS Goa.
I work at the intersection of deep learning and real-world systems — deploying vision pipelines on drones and embedded platforms, and researching 6-DoF spacecraft pose estimation and safe text-to-image diffusion models.
◇ SKILLS
- ▸ Python
- ▸ C/C++
- ▸ SQL
- ▸ PyTorch
- ▸ Scikit-Learn
- ▸ HuggingFace
- ▸ mmcv
- ▸ Docker
- ▸ Linux
- ▸ CUDA
- ▸ Git
- ▸ Jetson Orin
- ▸ Jetson Nano
- ▸ Raspberry Pi
- ▸ Arduino
- ▸ Pixhawk Cube
◇ EXPERIENCE
Deep Learning Researcher
May 2025 – Present- ▸ISRO-sponsored project on 6D spacecraft pose estimation using vision transformers.
- ▸Built T.A.R.S — a 100K-image synthetic spacecraft pose dataset across orbital + illumination conditions.
Research Intern
Mar 2026 – May 2026- ▸Research on diffusion-based T2I models, AI safety and decoding-time alignment.
- ▸Adversarial jailbreak attacks, RL control policies, localized inpainting for safe generation.
Robotics & Software Engineer
Aug 2024 – Nov 2025- ▸Deployed real-time CV pipelines on UAV/UGV embedded hardware with Docker + TensorRT.
- ▸Self-refining VLM trauma detection pipeline — ~80% accuracy on DARPA emergency dataset.
◇ PROJECTS
Llava-Lens
Logit-Lens interpretability for VLMs. Prompt-free transformer classifier over VLM logits — 83% on DARPA dataset.
ViT 6-DoF Pose
ViT-from-scratch on CIFAR-10 (76.5%). CLS-token 6D satellite pose via ViT-B transfer. Jetson Orin + TensorRT @ 17ms.
Wirefall
BERT-based web firewall trained on 2M safe HTTP requests · 98% accuracy. Novel KD-LoRA continual learning. 45ms inference.
◇ PUBLICATIONS
SPACE-HOP: Spacecraft 6-DoF Pose Estimation using Embedding-Predictive Pretraining and Hopf Map
↗ READJEPA pre-training for 6-DoF pose. Sub-degree accuracy via classification+refinement with discrete Hopf grid and continuous Lie offsets.
GuardPaint: Speculative Safety Decoding for Text-to-Image Generation
↗ READDecoding-time alignment combining timestep-aware localized inpainting with a REINFORCE diffusion policy over CFG, noise jitter and DDIM inversion.
◇ ACHIEVEMENTS
◇ CONTACT
Open to research collaborations and full-time opportunities. Reach out anytime.