Sharvesh Prabhu

SharveshPrabhu

I study computer science at VIT Vellore, specialising in data science. I'm in my second year.

This semester

Database Systems, Computer Networks and Exploratory Data Analysis

Working on

NexusTier

Middleware that moves old PostgreSQL data out to Parquet, keeps queries exact across both, and doesn't let analytics slow the transactions down.

See projects

About

I'm Sharvesh, a second-year B.Tech CSE student at VIT Vellore, specialising in data science.

Most of what I work on sits around data: cleaning it, exploring it, and figuring out how databases store and move it. Lately that has meant a lot of time inside PostgreSQL and DuckDB for my middleware project, and a lot of notebooks for EDA lab.

On the side, I'm slowly working through NeetCode problems in C++.

Away from the laptop it's mostly anime. Favourites so far: Bleach, Attack on Titan, Chainsaw Man and Sword Art Online.

This semester

  • Database Systems
  • Computer Networks
  • Exploratory Data Analysis
  • Probability and Statistics
  • Problem Solving using Java

Done

  • Data Structures and Algorithms
  • Operating Systems
  • Structured and Object-Oriented Programming
  • Discrete Mathematics and Linear Algebra

Projects

Research project

NexusTier

Transparent polyglot middleware for data tiering and HTAP

NexusTier sits in front of an unmodified PostgreSQL database. Closed monthly partitions of old data move out to compressed Parquet files that DuckDB queries, recent data stays in PostgreSQL, and queries still get exact answers across both.

A feedback controller watches transaction latency and decides how much analytics and data movement to allow, so the transactional side stays fast. I've benchmarked it on a TPC-C-derived workload running alongside the 22 CH-benCHmark analytical queries.

Built with Python / PostgreSQL / DuckDB / Parquet / Apache Arrow / FastAPI / sqlglot

View on GitHub ↗

How it fits together

Gateway
Plans each query across both tiers and aggregates partition by partition.
Controller
Watches transaction p99 latency and decides how much analytics and migration to let through.
Tiering
Moves closed monthly partitions to Parquet and never deletes the only copy of a row.

Datasets

  • Café sales (uncleaned)
  • Sales records
  • Wine Quality
  • Iris
  • Handwritten digits

Coursework

EDA lab work

Notebooks from my Exploratory Data Analysis course

Lab assessments from my EDA course at VIT: cleaning a messy café sales dataset, digging into distributions and correlations in the Wine Quality and Iris datasets, and visualising handwritten digits with t-SNE.

Built with Python / pandas / NumPy / Matplotlib / seaborn / scikit-learn / SciPy

View on GitHub ↗

Certifications

Passed

AI-200

Microsoft Certified: Azure AI Cloud Developer Associate

Exam AI-200 · Developing AI Cloud Solutions on Azure

Microsoft's associate-level exam on building, deploying, securing and monitoring AI solutions on Azure, written mostly in Python. It covers containerised services and working with Azure's data services.

Verify credential ↗

Contact

Email is the best way to reach me.