Home Experience Research Projects Education

Harsha Vardhan

A machine learning engineer passionate about building AI systems that solve real-world problems. Recently completed a Master's in Data Science from Rutgers University, where I focused on applied machine learning, backend systems, and end-to-end product development. Prior to that, I founded Autodub, a speech translation startup that was acqui-hired by Khabri (YC W19), where I worked as a founding ML engineer building large-scale audio intelligence tools. My interests lie in AI agents, foundation models, and the evolving role of intelligent systems in everyday workflows. When I'm not working, you'll probably find me exploring authentic restaurants around town or hunting down the next great meal.

Email / CV / Twitter / Github / LinkedIn

profile photo
Experience
LSDirect

Data Analyst Internship June 2024 - Dec 2024
LS Direct Marketing
Software: Pytorch, Python, SQL, VectorDB

Vedantu Learning

Machine Learning Engineer Nov 2021 - June 2023
Vedantu Learning
Software: Pytorch, Python, HF-Transformers, SQL, Tensorflow, Docker

CVSSP

Founding Engineer Dec 2020 - Nov 2021
Khabri Audio (YC W19)
Software: Pytorch, Python, HF-Transformers, SQL, Tensorflow, Docker

Autodub.in

Founder April 2018 - Dec 2020
Autodub
Software: Pytorch, Python, HF-Transformers, SQL, Tensorflow, Docker

Research
llm Capabalities Large Language Models: A survey on Use-case-Mapping & Evaluations
Zihan Xu, Harsha Vardhan, Mhamed Bettaieb,
Research ongoing — Paper in preparation
Projects
GensaveZ logo MR MOOLAH — PERSONAL FINANCE AI AGENT
Track your spending, compare with peers, and invest smarter with real-time, LLM-powered insights.
More details
  • Fine-tuned Llama 3.1-70B with LoRA adapters on AWS SageMaker distributed training (FSDP, 8 × A100 80 GB), cutting memory footprint by 65 % and improving domain-specific perplexity by 18 %.
  • Served the model behind a FastAPI gateway on a SageMaker Large-Model-Inference (Inf2.48xlarge) endpoint, delivering streaming responses (< 150 ms for 512 tokens).
  • Built a LangChain RAG agent that classifies transactions, pulls peer benchmarks from a Pinecone vector store, and generates personalized budgeting + asset-allocation advice.
  • Integrated Plaid OAuth to fetch live bank/credit-card data; implemented PII hashing and AWS KMS encryption-at-rest across PostgreSQL + S3.
Website
CSHORTS logo CSHORTS – AI NEWS SUMMARIZATION PLATFORM
Real-time, category-based news digests delivered in under 60 seconds.
More details
  • Fine-tuned & 8-bit-quantized a T5 model, slicing model size by 45 % while preserving ROUGE-L.
  • Deployed on AWS SageMaker with autoscaling & CloudWatch monitoring.
  • Built a FastAPI micro-service to ingest content, clean HTML, and store articles + summaries in MongoDB Atlas.
  • Streamlit front-end groups stories into Trending, Business, Tech & AI, Entertainment, Sports, USA, and India tabs; end-to-end latency < 60 s.
  • Containerised with Docker & CI/CD via GitHub Actions.
Code / Live demo
Ticket Triage System ENTERPRISE TICKET TRIAGE AUTOMATION SYSTEM
Data & Statistical Services Lab, Princeton University — funded by Chase Bank
AI-driven router that assigns and prioritizes support tickets with > 90 % accuracy.
More details
  • Fine-tuned MiniLM embeddings + CatBoost classifier to route tickets to the correct resolver teams and predict priority levels.
  • Achieved over 90 % macro-F1 on both classification tasks across a 1 M-ticket dataset.
  • Built a real-time FastAPI service that scores incoming tickets and publishes results to internal queues (Kafka).
  • Developed a lightweight Streamlit dashboard for Support Ops to search, monitor, and override triage decisions in-flight.
  • Implemented role-based access, audit logging, and encrypted data-at-rest to meet Chase's security requirements.
Education
NYU

MS in Data Science Sep 2023 - April 2024
Rutgers University

MEC

Bachelor of Technology (B.Tech) in CSE Aug 2016 - May 2020
Mahindra University Ecole Centrale School of Engineering

Updated on: June 11th, 2025