Building human-centered AI from research to reality.
I work on AI systems that understand people, motion, and the world around them — with a focus on 3D human pose, assistive computer vision, and trustworthy AI evaluation.
Creator of Vedansh Labs, an open-source research and engineering initiative for turning AI ideas into practical, reproducible systems.
Research on reliable evaluation of multi-hypothesis 3D human pose estimation.
I am studying the gap between the quality that exists inside a model's candidate set and the quality that a deployable selector can actually recover.
Technical focus
- Multi-hypothesis 3D pose estimation
- D3DP and MotionAGFormer-K
- MPI-INF-3DHP
- Distribution shift and selector reliability
- Leakage-controlled evaluation
- Reproducible experiment pipelines
Research Assistant, University of North Texas · NASA-funded project
SmartSight is a context-aware assistive vision system for blind and low-vision users. The goal is to move beyond basic object detection by reasoning about where an obstacle is, how it is moving, whether it blocks the user's path, and how urgently the user should be warned.
Current pipeline
Video / Camera
↓
YOLO Object Detection
↓
Spatial + Path Context
↓
Object Tracking
↓
Motion Reasoning
↓
Risk Scoring
↓
Adaptive Warning Generation
Technical focus
- YOLO object detection
- Spatial and path reasoning
- Temporal object tracking
- Approaching / receding / lateral motion classification
- Risk-aware warning generation
- Adaptive compute and alert gating
Vedansh Labs is my open-source AI research and engineering initiative focused on building practical systems around three themes:
| Area | Focus |
|---|---|
| Human Understanding | 3D human pose, motion, graph and transformer models |
| Accessible AI | Assistive computer vision and human-centered interfaces |
| Trustworthy Intelligence | Reliable evaluation, AI safety, and reproducible research tools |
The guiding idea is simple:
AI research should not stop at a paper or isolated experiment. It should become something people can inspect, reproduce, improve, and use.
LitVerify — Vedansh Labs
Faster, verifiable literature review.
A local-first workspace for organizing papers, evidence, contradictions, datasets, research gaps, and verifiable claims.
Built with: React · TypeScript · local-first storage · reproducible evidence workflows
LLM ShieldBench — Vedansh Labs
Lightweight evaluation infrastructure for safer conversational AI.
A benchmark platform for comparing LLM responses across prompts, models, and safety-oriented scoring criteria while keeping provider credentials non-persistent.
Built with: Python · React · LLM evaluation · safety benchmarking · reproducible result storage
Computer Vision
3D Human Pose Estimation
Human Motion Understanding
Graph Neural Networks
Transformers
PyTorch
Assistive / Accessibility AI
Embodied and Spatial AI
Trustworthy AI Evaluation
Reproducible ML Research
Languages: Python · C/C++ · JavaScript/TypeScript · SQL · Bash
ML / CV: PyTorch · Transformers · GNNs · YOLO · OpenCV · NumPy · scikit-learn
Research / Dev: Git/GitHub · Linux · CUDA · Docker · Jupyter · React · Node.js
I want to build AI systems that are not only capable, but also useful, reliable, reproducible, and centered around real human needs.
My long-term interests include:
- Human-centered spatial intelligence
- 3D human motion understanding
- Egocentric and embodied perception
- Assistive AI for blind and low-vision users
- Reliable evaluation of modern AI systems
- Open research infrastructure