Arnav Dhar

AI Engineer & Researcher · Durham, NC

CS (AI & ML) + Math at Duke. I research and build applied machine learning for health.

This summer I'm a software engineering intern at Gradient Health, where I built the evaluation harness behind the OCR engine in their medical-image de-identification system, and a research software engineering intern in the Hickey Lab, where I engineered VaultLab — an open-source Claude Code harness for end-to-end research. I also founded Anchor, an AI consultancy for luxury jewelry brands.

Download CVGet in touch

Education

  • Duke University

    Aug 2025 – May 2028

    B.S. Computer Science (AI & ML concentration) · B.S. Mathematics · GPA 3.89

    Duke Applied ML Director · Melissa & Doug Entrepreneurs Fellow · HackDuke 1st Place · Club Tennis

  • Northwood High School

    2021 – 2025

    Cum Laude · Irvine, CA · GPA 3.97 · SAT 1560

    National Merit Finalist · President's Volunteer Service Award (Gold) · ML Club President

Skills

Languages

  • Python · TypeScript · JavaScript
  • Java · C++ · SQL · HTML/CSS

ML & Data

  • PyTorch · TensorFlow
  • scikit-learn · XGBoost
  • NumPy · pandas · OpenCV
  • Computer vision · NLP · model evaluation

LLMs & Agents

  • LangGraph · multi-agent systems
  • RAG · vector search · embeddings
  • Prompt engineering
  • OCR

Web & Infrastructure

  • React · Next.js · Node.js
  • REST APIs · PostgreSQL
  • AWS · Google Cloud · Docker · BigQuery
  • Git · GitHub Actions · CI/CD

Experience & Projects

  1. Software Engineering Intern

    $20K/mo recovered

    Gradient Health · Durham, NC

    • Designed an evaluation harness to replace the OCR engine behind Gradient's medical-image de-identification system.
    • Built a reproducible 1,250-DICOM-image benchmark scoring 10+ OCR models (docTR, PaddleOCR, VLMs) on custom token-accuracy and false-redaction metrics, targeting the failure mode that corrupted clinical data.
    • Recovered $20K/month in sellable data by reducing OCR false redaction — an issue affecting 10% of customers and taking 4x a standard delivery to fix manually — under a HIPAA/self-hosting constraint.
    • OCR evals
    • Medical imaging
    • HIPAA

  2. Research Software Engineering Intern

    VaultLab

    Hickey Lab · Durham, NC

    • Engineered VaultLab, an open-source (MIT, published to PyPI) Claude Code harness that unifies literature search, data analysis, manuscript writing, and figure and slideshow generation into one LLM-readable knowledge base.
    • Built VaultLab's literature pipeline, fanning each query across 7 databases and ranking 200+ papers by conceptual relevance in a single batched LLM call.
    • Devised a citation-verification system that blocks manuscript export when any citation fails to match its source PDF, eliminating LLM hallucinations in generated research; used across 5 projects in the lab.
    • Multi-agent LLMs
    • Claude Code
    • Research automation

  3. Founder, Freelance Software Developer

    5 clients

    Anchor · Durham, NC

    • Founded Anchor, an AI consultancy for luxury jewelry brands, scaling to 5 paying clients including Ruchi NYC.
    • Automated 12 hours/week of labor with a full-stack lookbook, email, and wishlist automation suite.
    • Developed a cross-sell recommendation system with attribute filtering and metal normalization, using client-side caching for fast retrieval across 10k+ items.
    • Built catalog-intelligence infrastructure for Aura Jewels, a large jewelry manufacturer handling ~900 samples/month, including a visual similarity search that optimizes CAD design by preventing redundant sampling.
    • Luxury jewelry
    • Full-stack
    • Recommendations

  4. Harvest

    • Engineered a satellite-imagery pipeline delivering crop-health analysis in under a minute by streaming only the pixels inside a field boundary via HTTP range requests, instead of downloading full multi-gigabyte scenes.
    • Built a 5-node LangGraph agent pipeline on Claude fused with an interactive Next.js/Mapbox frontend, using satellite data, weather forecasts, and crop-specific thresholds to rank actionable farm recommendations with live-polling result updates.
    • LangGraph
    • Geospatial
    • Next.js

  5. Software Engineering Intern

    PROTECT3D · Durham, NC

    • Built a multi-stage cascaded XGBoost pipeline predicting custom orthotic design parameters from 3D foot-scan landmarks, at ~3% average error.
    • Replaced manual clinical measurement with automated ML inference; adopted across PROTECT3D's product line.
    • XGBoost
    • MedTech

  6. AI/ML Research Intern

    Mind Computing · San Diego, CA (Remote)

    • Improved clinical terminology translations for the Knowledge Platform, increasing AI model performance by 23%.
    • Built a speech-to-computable-text system for clinical encounters — runner-up in the national AI Tech Sprint among 152 companies.
    • Clinical NLP
    • Speech AI

Research

  • Tim Duong Research Laboratory

    Albert Einstein College of Medicine · AI Researcher · May 2024 – Aug 2025 · Remote

    • Developed AI to model the progression of Alzheimer's with 88% accuracy, supporting work published in PeerJ Computer Science.
    • Implemented a personalized breast-cancer treatment model with 83% accuracy as part of oncology AI research.
    • Created a standard operating procedure for accessing cloud VMs, now used by 150+ lab members.
    • LLMs
    • Predictive Modeling
    • Published

  • Cognitive Synergy Corporation

    U.S. Patent

    AI & ML Research Intern · Aug 2023 – May 2024 · San Diego, CA

  • RespRisk

    Published

    Independent Study · Nov 2024 – Aug 2025

Hobbies

Outside of work I'm usually on a tennis court, trying new food, or giving Coco belly rubs.

Elsewhere

EmailLinkedInGitHubCV

Reach me directly at arnav.dhar@duke.edu.