I build systems that connect visual perception to action: reconstructing spaces, detecting changes in 3D, and helping robots understand and interact with their surroundings.
I'm pursuing an M.S. in Computer Science, Machine Learning Track, at Columbia University, with graduation expected in December 2026. My work spans 3D vision, robot learning, and the engineering behind efficient AI systems.
Email · LinkedIn · Explore my repositories
Two phone recordings → a spatial report of what was added, removed, or moved.
A local pipeline combining joint COLMAP reconstruction, metric depth, occlusion-aware change detection, CLIP labeling, and interactive Rerun visualization. Includes the reasoning behind the design choices and the limits of monocular reconstruction.
Python · COLMAP · Depth Anything V2 · CLIP · Rerun
From image correspondences to camera poses and sparse geometry.
Built a structure-from-motion pipeline using SIFT, geometric filtering, RANSAC-based pose estimation, and iterative triangulation. Reported mean reprojection error below 0.8 pixels in project evaluation.
Computer vision · Multi-view geometry · Structure from motion
Natural-language instructions → object detection → pick-and-place control.
Built a 2D robotic manipulation system with a fine-tuned YOLOv11 detector and a UDP control pipeline sustaining 60 Hz servo updates, with 18 ms servo delays in project testing.
Robot learning · YOLOv11 · UDP · Real-time control
From real-world captures to interactive Gaussian-splatted objects in Unity.
Led a four-person Columbia team and architected an AnySplat-to-Unity pipeline for processing captures and streaming reconstructed assets through a Google OAuth-secured backend. The project supports object selection, runtime Gaussian-splat rendering, and 6DOF manipulation in a Meta Quest XR experience.
Unity · C# · AnySplat · Gaussian splatting · Meta Quest · OpenXR
Profiling and optimizing an agent workflow built on IBM's AssetOpsBench.
In a Columbia project, reduced workflow latency by 68%, accelerated model inference 3.3×, and reduced tool-call overhead 9× in test runs through direct batched inference, model preloading, and caching.
PyTorch Profiler · Weights & Biases · Inference optimization · AI agents
An interface for configuring, executing, and inspecting multi-step AI prompt chains.
Supports ordered chain editing, model selection, per-step outputs, timing and token records, and generation history. Built with Next.js and a shared backend, with server-side provider credentials and role-based access.
Next.js · LLM workflows · Full-stack development
- Columbia CAVE Lab · Research Assistant · Jan–Apr 2026 — Contributed multi-camera dataset collection and setup for Squash: 3D Object Assets and their Material Properties, and worked on ROS2 communication and tactile-feedback monitoring for object manipulation.
- IIT Madras Research Park, HTIC · Research Intern · Sep–Dec 2023 — Co-developed stereo-camera 3D mapping and integrated it with a modified SCARA robot for surgical registration; reported reconstruction error below 5 mm and latency below 100 ms.
- VIT · Research Intern · Mar–Dec 2024 — Designed a vision transformer for NIR-based root-rot detection, achieving 86% accuracy, and collaborated on a hydroponic growth chamber for training-data collection.
- Contributing author, “Beyond Static Leaderboards: Rank-Transfer Stability for the Evaluation of LLM Agents” — arXiv preprint, 2026.
- “Systems and Methods for Early Detection of Root Rot in Hydroponically Grown Plants” — patent application 202541009221, filed February 4, 2025.
- Languages: Python, C++, C#, TypeScript, JavaScript, SQL
- Perception & 3D: OpenCV, Open3D, COLMAP / pycolmap, Gaussian splatting, Agisoft Metashape, Rerun, Three.js
- Robotics & XR: ROS2, Unity, Unity XR Interaction Toolkit, OpenXR, Meta Quest, Unity Sentis, NVIDIA Omniverse, Arduino, UDP control
- ML & inference: PyTorch, TensorFlow, scikit-learn, Hugging Face Transformers, CUDA, vLLM
- Vision & action models: OpenVLA, YOLO, CLIP, Depth Anything V2 / V3, Qwen3-VL, AnySplat
- Agents & evaluation: LangChain, MCP / FastMCP, PyTorch Profiler, Weights & Biases
- Audio & transcription: librosa, faster-whisper, Qwen3-ASR, NVIDIA Parakeet, noisereduce
- Web & applications: React, Next.js, Tailwind CSS, Framer Motion, FastAPI, Streamlit
- Data & scientific computing: PostgreSQL, Supabase, NumPy, SciPy, Pandas, Matplotlib
- Development & deployment: Git, GitHub Actions, Docker, Linux, uv, AWS, SST, Vercel
I'm interested in spatially aware AI and robots that can move from interpreting a scene to acting within it. I want to deepen my work in robot learning, 3D scene understanding, and efficient perception-to-action systems, with an emphasis on reproducible evaluation and real-world constraints.
If you're working on related research or engineering problems, let's connect.
