Machine Learning • Full-Stack Engineering • Systems Reliability • Applied AI

---

About Me

Hey, I’m Arnav Kumar, a rising Junior at The University of Texas at Dallas building across machine learning, backend systems, cloud infrastructure, and full-stack applications.

I like working on projects that move past “cool demo” territory and into systems that actually behave like production software: structured data pipelines, resilient APIs, deployment workflows, model inference, observability, validation, and clean user-facing interfaces.

I’ve competed in 5+ hackathons, earned 2 wins, and have been building projects across AI/ML, sustainability, infrastructure reliability, computer vision, GPU tooling, and full-stack web apps.

---

What I Build

AI / ML Systems Model training and inference pipelines Fine-tuning and structured output validation ONNX export and CUDA-accelerated inference Computer vision and geospatial ML Applied AI for sustainability and safety

Full-Stack + Infrastructure TypeScript, React, Next.js, NestJS REST APIs, databases, and deployment workflows Kubernetes-native systems Observability with traces, metrics, and logs Cloud-ready backend architecture

---

Tech Stack

Languages

Frontend / Backend

Machine Learning / Data

Databases / Infrastructure

DevOps / Cloud / Observability

---

Featured Projects

---

Cascade: Kubernetes-Native Reliability & Chaos Engineering Platform

Tech: Kubernetes Docker Redpanda OpenTelemetry Prometheus Grafana FastAPI Python ClickHouse Qdrant Redis Chaos Mesh

Cascade is a reliability engineering platform designed to test, observe, and reason about distributed systems under failure.

It runs controlled chaos experiments against Kubernetes workloads, collects telemetry through traces, metrics, and logs, reconstructs incident timelines, and helps explain what broke, why it broke, and what should happen next.

Highlights

• Built a Kubernetes-native platform using kind, Chaos Mesh, and microservice targets.
• Implemented a telemetry/event backbone using Redpanda as a Kafka-compatible broker.
• Designed services for observation, stream enrichment, topology mapping, causal reconstruction, and incident timelines.
• Integrated observability through OpenTelemetry, Prometheus, and Grafana.
• Expanded architecture toward long-term memory using ClickHouse for telemetry storage and Qdrant for incident similarity search.
• Focused on safe chaos testing with blast-radius controls and protected-service safeguards.

---

GPUBoost: CUDA Workload Analyzer & Optimization Toolkit

Tech: Python PyTorch CUDA SQLite FastAPI React Recharts TailwindCSS GitHub Actions

GPUBoost is a performance analysis tool for understanding and optimizing GPU workloads, especially for CUDA-based PyTorch training runs.

The project focuses on profiling model training behavior, tracking run history, generating optimization recommendations, and making GPU bottlenecks easier to understand.

Highlights

• Built CUDA workload profiling around PyTorch and torch.profiler.
• Designed recommendations for mixed precision, batch size tuning, dataloader settings, pinned memory, and torch.compile.
• Added benchmark comparison, run history, validation reports, and structured output artifacts.
• Built toward a full-stack dashboard with FastAPI, React, Recharts, and TailwindCSS.
• Validated the system with a large automated test suite and production-style project structure.

---

TrailCheck: Full-Stack Trail Safety Intelligence Platform

Tech: NestJS Next.js TypeScript Prisma SQLite PostgreSQL QLoRA Pydantic ONNX Vercel Render

TrailCheck is an end-to-end trail safety application that combines structured environmental data, machine learning, backend orchestration, and a clean user interface.

The system analyzes trail conditions and produces structured safety classifications for outdoor decision-making.

Highlights

• Fine-tuned a 3B-parameter instruct model using QLoRA.
• Built an ML data pipeline from raw NOAA/NWS and NPS data.
• Generated 660 supervised fine-tuning examples across 12 hazard types and 4 severity levels.
• Designed backend AI orchestration with a local model path, Gemini fallback, and rule-based fallback.
• Used Pydantic validation to enforce a strict 5-field JSON response contract.
• Built the app with NestJS, Next.js, TypeScript, Prisma, and database-backed APIs.

---

Deforestation Detection Model: PyPI Package for Satellite Image Segmentation

Tech: Python ONNX Runtime CUDA U-Net Sentinel-2 Google Earth Engine PyPI

A computer vision package for detecting deforestation from 10-band Sentinel-2 satellite imagery.

The model performs semantic segmentation over geospatial tiles and is packaged as an installable CLI tool so it can be used outside of a notebook or research environment.

Highlights

• Trained a U-Net segmentation model on Sentinel-2 satellite imagery.
• Used labeled forest-loss data from University of Maryland datasets and Google Earth Engine workflows.
• Exported model inference to ONNX Runtime with optional CUDA acceleration.
• Achieved fast tile-level inference on 512x512 imagery.
• Packaged the project as a PyPI-installable CLI with CPU/GPU support.
• Designed class outputs for non-forest, forest, and deforested regions.

---

Operations Task Tracker: AWS-Oriented Full-Stack CRUD App

Tech: React Node.js Express JavaScript AWS REST APIs

A full-stack task management application built to practice production-style CRUD flows and cloud deployment patterns.

Highlights

• Built task creation, editing, deletion, assignment, priority tracking, and status updates.
• Designed REST API routes with a Node/Express backend.
• Connected the frontend and backend through environment-based API configuration.
• Explored AWS deployment paths for frontend and backend services.

---

Current Areas of Interest

Applied AI Systems       → ML models that solve real operational problems
Reliability Engineering  → chaos testing, observability, incident reconstruction
GPU Optimization         → CUDA profiling, training performance, ML systems tooling
Sustainability AI        → profitable environmental intelligence for businesses
Computer Vision          → satellite imagery, segmentation, geospatial ML
Full-Stack Products      → clean frontends backed by serious backend architecture