Accenture · Machine Learning Engineer Intern
Worked on computer vision and anomaly detection for drone footage, connecting ML models to industrial inspection and real-world data pipelines.
UC Berkeley EECS Graduate · Machine Learning · Autonomy
Building technical systems at the edge of software, AI, and the physical world.
I’m Aadit, a UC Berkeley EECS graduate interested in machine learning, software engineering, and systems that connect models to real-world environments. My work spans computer vision, perception/SLAM for autonomous systems, reinforcement learning, and hardware-adjacent software.
Experience
Where I’ve worked
Amazon · Software Development Engineer Intern
Built database and infrastructure systems with a focus on reliability, backend engineering, and scalable production software.
Formula Electric Berkeley · Perception Lead
Led perception work for autonomous systems, with emphasis on robotics, mapping, and software that sits close to hardware.
Blume Venture Capital · Analyst
Supported investment analysis and portfolio strategy, leveraging market insights and financial acumen to drive data-backed decisions.
Convertible AI · Machine Learning Engineer
Worked on recommendation systems at an early-stage AI startup, combining product iteration with applied ML engineering.
Research, Papers, & Projects
Recent technical work
Selected research areas, papers, projects, and implementation studies spanning applied ML, computer vision, autonomy, and language-model alignment.
Optimization
AdamW Transferability
Research on how AdamW optimizer behavior transfers across model scales and architectures, with emphasis on generalization and training dynamics.
Paper coming soon →Language-Model Alignment
Confidence-Weighted Preference Optimization for Instruction-Following RLHF
Compared offline preference optimization, reward modeling, and online RLHF methods under a controlled instruction-following evaluation. The study evaluates DPO, IPO, AOT-style quantile matching, Bradley-Terry reward modeling, GRPO, DrGRPO, GSPO, and a confidence-weighted DPO variant that upweights clearer reference-corrected preference margins.
Read paper →
Remote Sensing
UAV-Based Tree Crown Detection for Urban Tree Stress Analysis
Developed a multimodal vision pipeline for detecting and segmenting individual tree crowns from UAV imagery, combining RGB, vegetation, and height-derived signals with GroundingDINO, Segment Anything, and bounding-box refinement.
AI Systems
KaRL
A coordination layer for companies using multiple reinforcement-learning-powered AI agents. The project explores how independently learning agents can interfere with each other and how coordination infrastructure can improve multi-agent behavior.
Discuss project →Computational Imaging
Light Field Refocusing and Synthetic Aperture Rendering
Implemented light-field rendering techniques for computational refocusing and synthetic aperture control using structured camera arrays and handheld capture experiments.
Augmented Reality
Marker-Based Augmented Reality with Camera Calibration
Built a camera-calibrated AR pipeline that tracks 2D keypoints, estimates projection matrices, and renders 3D geometry into real video frames.
Diffusion Models
Denoising U-Net Training for Digit Diffusion
Trained a compact U-Net diffusion model for noisy image restoration, timestep-conditioned denoising, and label-conditioned digit generation.
Generative AI
Diffusion-Based Image Generation, Editing, and Restoration
Explored diffusion-based generation and editing with iterative denoising, classifier-free guidance, image-to-image translation, inpainting, and prompt-conditioned visual transformations.
Robust Estimation
Panorama Reconstruction with Homographies, Feature Matching, and RANSAC
Built a panorama reconstruction pipeline combining homography estimation, feature detection, adaptive non-maximal suppression, descriptor matching, RANSAC, and image warping.
Geometric Vision
Geometric Face Morphing and Population-Average Modeling
Developed a geometry-based morphing pipeline using facial correspondences, Delaunay triangulation, affine warping, population averages, and controlled extrapolation.
Frequency Analysis
Frequency-Domain Image Processing and Multi-Resolution Blending
Implemented classical frequency-domain and multi-resolution techniques for edge detection, sharpening, hybrid images, Fourier analysis, and seamless image blending.
Computational Imaging
Computational Restoration of Historical Glass-Plate Photography
Built a multiscale image-alignment pipeline to reconstruct color photographs from historical glass-plate negatives using pyramid search, edge-aware scoring, and channel registration.
Contact
Interested in AI, autonomy, physical AI, startups, or research?
I’m always happy to talk about technical work, early-stage ideas, and opportunities at the intersection of machine learning and real-world systems.