cv

Research Engineering Intern

HeyGen

Los Angeles, California

• Fine-tuned a camera-motion ControlNet using point-cloud renders and pose inputs to guide multi-modal Video Diffusion Models (Wan 2.1), producing smooth, spatially consistent camera trajectories for AR content.
• Leveraged sparse background optical flow and early fusion to develop a multi-modal classifier to classify camera motions as part of the internal data filtration pipeline - increasing the pass rate by 10% and data purity by 20%.
• Built and parallelized a large-scale SfM pipeline (using VGGSfM and dynamic scene segmentation) to extract and align camera poses from 100K+ in-the-wild videos.

Graduate Research Assistant

Human Sensing Laboratory | Meta Reality Labs

Advisor: Prof. Fernando De la Torre · Pittsburgh, PA

• Exploring how 2D VLMs gain 3D scene understanding through multi-view reasoning and view selection, integrating Gaussian Splatting + depth representations, and analyzing the emergence of this 3D understanding via mechanistic interpretability.
• Developed an adversarial scene exploration framework combining SE(3) manifold optimization and ordinal-loss functions to expose failure modes in depth estimation and geometry reconstruction (CVPR 2026, under review).

Machine Learning Engineer

Arintra

Bangalore, India

• Improved ICD code prediction accuracy by 6% via Retrieval-Augmented Generation (RAG) with medical LLMs leveraging disease comorbidity knowledge.
• Designed a semantic search engine using SapBERT + Qdrant vector DB for medication retrieval, boosting F1 by 11% and generalizability across 4 hospitals.
• Developed a model versioning and deployment system using MLFlow, FastAPI, and GCP, streamlining engineer workflows, cutting model deployment time by ~20-25%, and ensuring reproducible, traceable releases of finetuned models.

Machine Learning Engineer

Klothed

Advisor: Prof. James O'Brien · New York, USA

• Enhanced texture fidelity for single-view 3D human mesh reconstruction (ECON) using diffusion-based texture synthesis, image super-resolution and synthetic data augmentations.
• Accelerated a 2D image warping pipeline by 90% (2s → 0.2s) through finite-element optimization, enabling real-time inference for AR applications.
• Proposed a synthetic data pipeline in Blender to render diverse clothing conditions (lighting, draping), resulting in ~2% improvement in MSE for image matting models.

Research Engineer

Origin Health

Advisor: Dr. Sripad Devalla · Raffles Quay, Singapore

• Co-authored a paper combining domain-specific synthetic data and a novel heatmap-based attention mechanism to achieve a 3.8% reduction in MAE for fetal biometry measurements compared to SoTA approaches (IEEE ISBI 2022).

MSCV

Carnegie Mellon University — Robotics Institute

Master of Science in Computer Vision

• GPA: 4.11/4.0

Bachelor of Technology

Manipal Institute of Technology

Electronics and Communication

• GPA: 8.92/10

Breaking Depth Estimation Models with Semantic Adversarial Attacks

CVPR 2026 (Under Review)

Narayan, A.*, Ojha, U*., Theiss, J., Prakash, A., De la Torre, F.

Device-Independent Deep Learning System for the Automated Segmentation of Sonographic Fetal Brain Structures

SPIE Medical Imaging 2022

Multi-center, multi-device validation for fetal brain structure segmentation.

Biometric Constraint Attention Masks to Obtain Sonographic Measurements

IEEE ISBI 2022

Co-authored paper on attention mechanisms for fetal biometry measurements.