Start with the real problem
Understand the task, user context, and product constraints before deciding what the model should do.
Profile / Research and systems
About Harsha
I am an Applied Research Scientist who works across machine learning research, engineering, and product constraints.
I started in computer vision and robotics, moved through security and consumer software, and now focus on language models, agentic systems, multimodal ML, and the people those systems are meant to help.
01 / Approach
How I approach applied research
Good model work is necessary. A useful system also needs the right problem, evidence, and constraints.
Build an evidence loop
Make quality, latency, failure modes, and human outcomes visible early enough to change the design.
Carry the work through
Move between research, engineering, and product details until the system holds up outside a demo.
02 / Career
A path through research and engineering
From 3D reconstruction and robotics to real-time inference and production ML.
Amazon
$200M+ IMPACT / MULTI-AGENT SYSTEMS
Senior Applied Scientist
Leading science strategy and architecture across Seller Experience and Seller Agentic Intelligence. The work spans multi-agent systems, tool use, model post-training, generative UI, and automation initiatives that have delivered more than $200M in operational savings.
Akamai / Ghost Security
20x LATENCY REDUCTION
Machine Learning Intern
Applied ML to web application firewall inference and reduced per-request latency from 1 millisecond to 50 microseconds.
UC San Diego
VISION / ROBOTICS / AUTONOMY
Graduate Research
Researched computer vision and robotics across autonomous systems and health applications, including motion understanding from moving camera platforms.
Samsung R&D
40% SMALLER / 30% FASTER
Software Engineer
Refactored the 150K-line S-Planner codebase, reducing its size by 40% and launch time by 30%. Also built calendar features and worked on early foldable-screen prototypes.
NIT Surat
3D VISION / CUDA
B.Tech and Computer Vision Research
Worked on 3D reconstruction from Kinect and RGB cameras and used CUDA to accelerate the underlying computer vision workloads.
03 / Beyond work
Other systems I keep taking apart
The subject changes, but the habit is usually the same: inspect it, understand it, then build something.
Photography
Night skylines, landscapes, and basic astrophotography. See selected work on 500px.
Physical computing
Custom PCs, 3D printing, home automation, and the practical consequences of airflow and tolerances.
Reading and curation
Research papers, systems writing, and the curation pipeline behind Paper Radar.