Aarush Narang

I build  

I'm drawn to problems where being right matters: a near-miss at an intersection, a finding on a chest X-ray, a ticket someone's been waiting on since morning. I'm a data scientist and ML engineer with two years shipping production ML at HCL Technologies, a paper at DEXA in Naples, and a habit of building things nobody asked me to. Now an MS in Data Science at the University of Arizona, and Research & Development at Biosphere 2.

View My Work View Resume

About Me

Aarush Narang

Most of what I've learned sits in the gap between a model that works in a notebook and one people depend on every day. I spent two years at HCL Technologies in that gap, building Gen AI copilots that read across 4,000+ enterprise documents and cost models that run as a weekly job on Azure, all of it behind Docker and CI/CD, because a model nobody can deploy isn't finished. Before that came Carrier and PwC, where I shipped an FAQ bot that took roughly a third of the load off a support inbox and a Neo4j graph untangling how cloud-native projects depend on one another.

I'm midway through an MS in Data Science at the University of Arizona (2025–2027, GPA 3.83), after a B.Tech in Computer Science from Shiv Nadar University and a paper at DEXA 2024 in Naples. The projects I pick up on my own keep circling the same territory, traffic footage and chest X-rays, where a missed call carries a cost and getting the evaluation honest matters more than the leaderboard.

Technical Skills

Languages
PythonRTypeScriptJavaScriptC#C++JavaGolangSQL
Machine Learning
PyTorchPyTorch LightningTensorFlowscikit-learnXGBoostLightGBMCatBoostSHAPpandasNumPyMONAItimmYOLO11OpenCVGrad-CAMONNXTensorRTWeights & BiasesLangChainHugging FaceRAGLLMs
Backend & Web
FastAPI.NETReactNode.jsREST APIsGradioHTMLCSSTailwind CSS
Databases
PostgreSQLMS SQL ServerAzure SQLMySQLMongoDBNeo4jCassandraHBaseOracle DatabaseSQLiteJDBC
Cloud & DevOps
AWSAmazon RDSECS FargateECRSecrets ManagerCloudWatchAzure App ServiceAzure DevOpsGCPCloud RunDockerGitHub ActionsCI/CDPrometheusGit

Experience

Eller Immersion Project ·

Aug 2026 – Present

Tiered Minute Reduction (TMR) Forecasting

  • Building a TMR forecasting model on Amazon-provided synthetic data that predicts weather-driven delivery capacity loss at the region level 5 days ahead, capturing day-over-day snow accumulation that the current planning process treats as independent 24-hour windows buffered network-wide.
  • Developing and comparing two model architectures, a distributed-lag logistic regression and a gradient-boosting ensemble (XGBoost, LightGBM, CatBoost) with monotonic constraints and SHAP feature attribution, evaluated by walk-forward backtesting grouped by storm event and selected on calibration and accuracy lift at the 5-day horizon.
  • Designing a simulation of the Return to Station (RTS) re-induction cycle to test hold policies that keep undelivered packages from inflating next-day demand during multi-day weather events.
XGBoostLightGBMCatBoostSHAP

AI/ML & Software Engineer ·

Aug 2026 – Present

GreenThumb Cloud Control System, Freight Farms

  • Built a sensor data-quality validation layer with Oracle SQL over the GreenThumb Cloud Control System's 26-table schema, classifying 58 sensors' readings against agronomic set points and dead-band ranges to flag values the inherited pipeline missed, testing locally via Docker Compose (Oracle 21c XE, Spring Boot, Python simulator) against 10,000+ real readings before validating against the live Amazon RDS Oracle schema via DBeaver and JDBC.
  • Secured edge access to the in-farm Raspberry Pi over Tailscale VPN and SSH, tracing the telemetry path (SSE ingestion, SQLite buffering, JSON POST to a Spring Boot backend on AWS ECS Fargate) and mapping the sensor-to-database schema.
OracleDockerSpring BootAmazon RDSRaspberry PiTailscale

Graduate Grader ·

Spring 2026

Computational Modeling in Cognitive Science

  • Hands-on MATLAB course on modeling learning and decision-making, and linking models to behavioral and neural data.

Software Engineer ·

Jul 2024 – Jul 2025

Generative AI Development Team

  • Built custom copilots (Python, TypeScript, LangChain, Hugging Face) integrated with ServiceNow and SharePoint for contextual search across 4,000+ knowledge files, reducing daily service desk tickets 26% (2,300→1,700) and resolution time 76% (105→25 min), shipped via CI/CD with automated tests.
  • Developed and deployed employee cost models over SharePoint data (Python, pandas, FastAPI, Azure SQL), with predicted cost landing within 7% of actual on average (MAPE), running as a semi-automated weekly workflow on Azure App Service, containerized with Docker and released through Azure DevOps CI/CD.
PythonTypeScriptLangChainHugging FaceFastAPIAzureDocker

Software Engineer Intern ·

Jan 2024 – Jul 2024

Sales KPP Enterprise Application

  • Developed the Sales KPP enterprise application end to end with React, C#, .NET & MS SQL Server, tracking employee sales targets and achievements against revenue and annual deal closure metrics, replacing manual Excel consolidation for HR, PMO, and L2 leadership.
  • Deployed the application to production and delivered the technical documentation.
ReactC#.NETMS SQL Server

Web Developer Intern ·

May 2023 – Jul 2023
  • Designed and developed the Account Payable Management System and Ticket Tracker for the sales team using HTML, CSS, JavaScript, jQuery & jQWidgets.
  • Integrated an FAQ chatbot that cut support queries 38% (450→280 per week), and added automated tests.
JavaScriptjQueryjQWidgetsHTMLCSS

Technical Intern ·

May 2022 – Sep 2022
  • Extracted and parsed project metadata from CNCF repositories with Python and REST APIs, then built a graph database model in Neo4j mapping dependency relationships across cloud native projects.
  • Correlated and visualized the data in Golang-backed dashboards running in Docker, surfacing insights for the Advisory team.
PythonNeo4jGolangREST APIsDocker

Featured Projects

Reflex detecting and tracking road users at an intersection, with risk lines between vehicles and pedestrians

Reflex — Collision Risk Detection

94.4% vehicle detection F1, validated against the Urban Tracker benchmark

Scans traffic video and flags dangerous close calls between cars, cyclists, and pedestrians. Uses YOLO11, ByteTrack, and OpenCV to detect and track road users, flags near misses at 87.6% precision against hand labeled events, and is optimized with TensorRT to cut processing time 3.8x (12→46 fps) for live feeds. Ships a dashboard that replays flagged events and maps where close calls cluster to reveal a city's riskiest intersections.

PythonPyTorchYOLO11OpenCVFastAPIReactPostgreSQL
Code
Chest X-ray from the NIH ChestX-ray14 dataset used by RadarMD

RadarMD — Chest X-ray Triage

0.84 mean AUROC across 14 thoracic pathologies — a 6 point gain over baseline

Deep learning triage system for chest X-rays, built on NIH ChestX-ray14 (112,000+ images, 30,000 patients). Trains DenseNet-121 and ConvNeXt with PyTorch Lightning under strict patient-level splits, reaching 0.84 mean AUROC across 60+ Weights & Biases experiments. Grad-CAM localization validated against 880 ground truth bounding boxes, with a test suite prioritizing serious findings over raw accuracy to keep missed diagnoses under 8%. Deployed via ONNX export for 4x faster CPU inference (~120 ms/image), serving a Gradio app on Cloud Run with GitHub Actions CI/CD and Prometheus metrics.

PyTorchMONAItimmGrad-CAMONNXFastAPIDockerGCP
Code
Human parsing model output

Self-Correction for Human Parsing

84.7% mIoU on LIP — 2.3% above the published SCHP baseline

Benchmarked the SCHP human parsing model against state-of-the-art methods across Look into Person (LIP), Active Template Regression, and Pascal Part, addressing label noise challenges.

PythonPyTorchComputer Vision
Code
Publication · DEXA 2024 · Naples, Italy

Analyzing the Efficacy of Large Language Models: A Comparative Study

Peer-reviewed paper (co-authored) presented at the 35th International Conference on Database and Expert Systems Applications, published in the Springer LNCS series.

Read Paper

Let's Collaborate!