I'm Arnav Kumar, a rising Junior at UT Dallas, currently a Data Engineering Intern at Resilient Inc. Previously an AI/ML Engineering Intern at Flexon Technologies, where I helped ship an NLP-driven ETL pipeline to production.
My focus is data engineering: pipelines, orchestration, and the agentic systems that keep them healthy. I build things that hold up under real constraints: deterministic safety gates instead of "trust the LLM," bias audits instead of "assume it's fair," measured results instead of assumed ones. Two of the projects below (below in Featured Projects) are ones I can walk through in depth, decision by decision, in an interview.
5+ hackathons, 2 wins.
Data Engineering Intern at Resilient Inc. · Current
AI/ML Engineering Intern at Flexon Technologies
- Built and deployed an NLP-driven ETL pipeline to production, processing documents at 130/minute, including a full run across 600 resumes.
Project Manager at IEEE Computational Imagery, UTD
- Coordinated project planning and cross-functional execution for the chapter's computational imaging initiatives.
Front-End Developer at Project Management Club (PMC), UTD
- Migrated the club site from vanilla HTML/CSS to Next.js, improving load speed by ~200ms, verified with Google PageSpeed Insights.
|
|
|
|
Languages
Data Engineering & Orchestration
AI / Agentic Systems
Cloud, DevOps & Full-Stack
Multi-agent incident response for data pipelines: detects, diagnoses, and remediates schema-drift incidents, with a hard stop between "the model suggests it" and "the system does it."
Tech: Python LangGraph Redpanda Kubernetes Ollama (Qwen3) NetworkX JSON Schema
Six LangGraph agents split the job cleanly: Detector and Diagnosis figure out what broke and why, Remediation-Planner proposes a fix, Safety-Checker decides whether it's safe to act, Executor runs it, Reporter writes it up.
The one decision I'd defend hardest: the Safety-Checker is plain deterministic Python, not an LLM call. An agentic system that lets a language model both diagnose and approve its own fix is a system waiting to auto-execute something wrong with total confidence. Instead, every action is scored for blast radius against a hard ceiling (MAX_AUTO_APPROVE_BLAST_RADIUS=15); anything over that gets denied and logged to escalations.log instead of auto-executing.
- Built and verified in phases, with the test suite growing as real components replaced stubs: 10/10 → 23/23 → 35/35 passing.
- Verified live against a real Redpanda cluster, not mocks, including 9-node blast-radius calculation via a NetworkX dependency graph.
- The Executor runs against a real local Kubernetes (
kind) cluster with a genuine rollback path, not a simulated one. - Local LLM (Qwen3 via Ollama) handles diagnosis reasoning and postmortem write-ups, no cloud API dependency.
- MIT-licensed, pushed to GitHub.
Graph-based fraud-ring detection for anti-money-laundering (AML) compliance, built to hold up under the same scrutiny real financial infrastructure gets.
Tech: Rust PyTorch Geometric LangGraph Azure Cosmos DB (Gremlin) Azure Event Hubs Tableau Azure
A Rust ingestion engine feeds transaction data into a graph, where a PyTorch Geometric GNN catches fraud rings that look clean at the individual-transaction level but form an obvious pattern once you look at the relationships between accounts. A LangGraph multi-agent pipeline then drafts Suspicious Activity Reports (SARs) for compliance review.
What made this a real engineering exercise rather than a demo were the bugs I found and had to actually fix:
-
An AMQP batching defect in the Event Hubs ingestion path that was silently dropping/misordering messages under load.
-
A Cosmos DB partition-key design flaw: the original key choice created hot partitions; corrected the architecture rather than patching around it.
-
An empty-output guardrail gap in the SAR generation pipeline that could let a blank report pass silently.
-
Access is locked down with Entra ID RBAC on the Gremlin API, no open graph access.
-
Tagged
v1.0with a 31-test CI suite and zero credential leaks across 28 commits. -
Compliance metrics visualized in Tableau.
| Project | What it does | Result |
|---|---|---|
| ViewFinder | Local, offline CLI that predicts geolocation from an image/video (GeoCLIP + a custom OpenCLIP/FAISS retrieval system) with a self-built Airflow ETL pipeline for reference data | Fine-tuning cut median geolocation error by up to 68% |
| Bench Redeployment Optimizer | 7-agent LangGraph system for IT-staffing bench management, governed by four-fifths-rule bias audits and human-in-the-loop gating | 222 passing tests, full audit-log coverage |
| Cascade | Kubernetes-native chaos engineering platform, 23 microservices, blast-radius scoring, full OTel/Prometheus/Grafana observability | 177-test suite |
| GPUBoost | CUDA/PyTorch GPU profiling & optimization CLI, published to PyPI (Repo →) | 3.2x training throughput improvement |
| TrailCheck | Full-stack trail safety platform (NestJS, Next.js, Prisma) backed by a QLoRA-fine-tuned 3B model | 660 hand-built SFT examples across 12 hazard types |
| Deforestation Detection | U-Net satellite image segmentation, published to PyPI with weights on Hugging Face | Fast tile-level inference on 512×512 imagery |
- 🏆 HackSMU VII: Award winner, iMasons Infrastructure track
- 🏆 RowdyHacks: Best Use of ElevenLabs
Data Engineering & Governance → orchestrated pipelines, agentic QA, lineage & compliance auditing
Applied AI Systems → multi-agent systems (LangGraph) for real operational problems
Reliability Engineering → chaos testing, observability, incident reconstruction
Computer Vision & Geospatial → geolocation, satellite imagery, segmentation
Full-Stack Products → clean frontends backed by serious backend architecture