I'm Aradhya, a builder and AI technologist from India. I specialize in blending embedded systems with AI—creating agentic applications that bridge the physical and digital worlds.

From engineering runtime memory trackers for autonomous vehicles at NVIDIA to building multimodal drone command frameworks with LLMs, my work is driven by obsession and precision.

Outside tech, I enjoy exploring new ideas, mentoring, and spending time with music and machines.

My work ethic is research centric, optimizing everytime for best possible outcomes.

X / Twitter LinkedIn Instagram Github

Experience

Project Intern

NVIDIA

Feb 2025 – Present

Developed a C++ memory tracking tool for autonomous vehicle software.
Tracked 100% of modern C++ allocations on ARM-based SBCs.
Integrated the tool with ROS-based autopilot stacks for runtime validation.

Software Development Engineer (Avionics)

AEROKLE

Nov 2022 - Oct 2023

Improved UAV navigation accuracy to ~0.5 feet
Done by implementing Ground Sampling Distance based triangulation for precise GPS localization.
Helped my team win All India Rank 7 in Aerothon 2024 (Phase 1) and All India Rank 2 in ADDC 2023

Tech Stack

Programming Languages

CC++PythonJavaScriptBash

Other Languages

SQLHTMLCSS

Frameworks/Technologies

React.jsNode.jsExpress.jsPyTorchOpenCVscikit-learnLangChain

Currently Exploring

QNX OSGDBAPI DevelopmentCore CodingMongoDBNext.jsTailwind CSS

Education

KLE Technological University

B.E. in Computer Science and AI 2022–2026

CGPA: 8.15 / 10
Solved 150+ DSA Questions on Leetcode and Codeforces
Published three research papers in Springer's LNCS

eStellar Academy

CBSE Higher Secondary Education 2020–2022

Achieved AIR 8115 in COMEDK 2022
Scored 88% in class 12th CBSE Examinations

Mahila & Nutan English Medium School

Goa Board SSC 2016–2020

Elected as Head-Boy with a 200+ vote margin
Scored 86% in class 10th GBHSE Examinations

My Work

Some projects I worked on, Obsessively of course

Computer Vision

PolluxPenguin: Agentic Drone Control

Built an end-to-end agentic system for ArduCopter using LLMs and RAG pipelines. Enabled drones to interpret natural language commands with multimodal feedback and contextual navigation.

DroneKitLLMROSReact.jsNode.js

Achieved contextual control accuracy of ~80% in simulation.
Tested on Gazebo with real-time command parsing.

GitHub YouTube

ML Forecasting

KANKAN: Time Series Prediction

Developed a Kolmogorov–Arnold Network (KAN) model with extremely low compute cost, delivering MLP-level accuracy with 99% fewer FLOPs for time series forecasting.

PyKANKANSeabornMatplotlib

Accepted for publication for Springer's LNCS by ICCCT 2025.
Visualized predictions and compared performance against MLPs.

GitHub Accepted Paper

NLP + Database

QueryGenie: SQL via RAG

A dynamic SQL query generator that uses RAG and LLMs to interpret user prompts and produce schema-aware SQL queries for any connected database.

LangChainPineconeNode.jsSQL

Adapts to custom schemas using a vectorized knowledge base.
Delivers accurate production-grade SQL with fallback prompts.

GitHub

HealthTech / Edge AI

Multi-Modal Anemia Detection System

Built a real-time embedded intelligence system for non-invasive anemia screening using smartphone images of the fingernail, palm, and conjunctiva. The mobile application operates fully offline to ensure privacy and rapid diagnostics in resource-constrained environments.

MobileNet V2TensorFlow LiteCNNAndroid

Achieved a peak accuracy of 88.6% and an AUC of 0.91 using the conjunctiva-based model.
Optimized quantized models to execute on-device inference in under 14 milliseconds on standard Android hardware.

Paper GitHub

Natural Language Processing

LCM-MT: Concept-Grounded Machine Translation

Designed a novel concept-based machine translation framework using Large Concept Models (LCMs) for English-to-Hindi translation. Utilized the Meta AI SONAR encoder to generate language-agnostic embeddings, mapping them to a Transformer decoder via an adapter layer.

SONAR EncoderTransformer DecoderAdapter LayerNLP

Recorded a BLEU-1 score of 0.9736 and a BLEU-4 score of 0.4605 using the IIT-Bombay Parallel Corpus.
Outperformed a baseline token-level Transformer model by achieving a +0.12 gain in the BLEU-4 metric.

Paper