Utkarsh Sonawane

I study computer science at Penn State. I’m fascinated by how large language models (LLMs) work under the hood — from transformers and inference to agentic AI systems — and I like turning that curiosity into software people can actually use.

Right now, I’m doing undergraduate research with Dr. Md Sultan Al Nahian on using LLMs to extract adverse drug reactions from Reddit. We’re exploring retrieval-augmented generation (RAG), agentic verification, and evidence-backed extraction to make these systems more reliable.

In my free time, I love doomscrolling X, working out while vibing to Chris Luno, and reading books.

Looking for Summer 2027 software engineering and AI/ML internships.

Computer science · Penn State · Class of 2028

Updates

Selected projects

Things I’ve built, and what I learned from them.

Nytr

A nutrition and training app built around Penn State dining data. Deterministic recommendations, a Python backend, and a native iOS client.

lessons from building it

Python, FastAPI, PostgreSQL (Supabase, row-level security) · Swift, SwiftUI, HealthKit

Case study Code Demo
Technical details

A nutrition and training app built around the food my campus dining hall serves each day. A Python backend computes meal and training guidance with exact arithmetic and versioned, append-only records. A SwiftUI app reads HealthKit and shows the result.

1,135 backend tests (1,124 passing in CI with Postgres) · 251 iOS XCTest cases

Clage

A from-scratch neuroevolution engine and local browser workbench for inspecting agents, experiments, and deterministic replay.

Python, FastAPI, JavaScript, Canvas · pytest, Playwright, mypy, Ruff

Case study Code Demo
Technical details

From-scratch NEAT engine evolving neural agents in a 2D ecosystem. Clage Studio 1.0 adds live browser visualization, neural-network inspection, evolutionary lineage, reproducible experiments, and deterministic replay.

429 Python tests + 4 JavaScript tests + 21 browser workflows in CI · OR/AND/XOR 5/5; sine 0/5 with minimal initialization

Talks

Record on Apple Watch; transcribe and summarize on iPhone; save to Notion. Built around recoverable transfers and persistent jobs.

Swift 6, SwiftUI, watchOS, iOS · WatchConnectivity · SpeechAnalyzer · Foundation Models · Notion API

Case study Code
Technical details

Record a meeting on Apple Watch. The audio moves to the iPhone in the background, is transcribed and summarized on device, and is saved to Notion. A persistent job queue and acknowledgement-gated deletion keep recordings safe through crashes and disconnects.

57 XCTest cases passing in CI · iOS and watchOS builds on macOS runners

Devvy

One local daemon coordinates Discord presence across coding tools while keeping prompts, code, and full paths private.

JavaScript (Node.js), TypeScript integrations · Discord IPC · GitHub Actions

Case study Code
Technical details

A local daemon that owns one Discord connection and picks which coding tool to show. It speaks Discord's binary IPC protocol directly and never shows prompts, code, or paths.

26 test scripts, all run in CI · releases only after CI passes · macOS released

More experiments

Research

Extracting adverse drug reaction reports from Reddit

Undergraduate research · Penn State · 2026–present

In progress

With Dr. Md Sultan Al Nahian, I’m studying how language models can identify reported drug–reaction relationships in Reddit discussions. The work compares direct prompting, retrieval-augmented generation, and extraction with verification against human annotations.

Evaluation focuses on precision, recall, F1, and error analysis.

Read the research notebook →

Research plan

As an Undergraduate Research Assistant at Penn State (REU), I am investigating LLM-based detection of adverse drug reaction reports and extraction of suspected drug–reaction relationships. The research plan compares direct prompting, retrieval-augmented generation (RAG), and agent-based extraction and verification. Planned analysis includes normalizing drug and reaction terminology and comparing extracted relationships with PubMed literature.

Elsewhere

Undergraduate Research Assistant → Research

Technical Chair, IEEE Student Branch, Penn State Harrisburg

Sept 2026 – Present

Elected in Sept 2026 after a year on the technical team (Sept 2025 – Sept 2026). Coordinate the branch's software, AI, and robotics workshops and projects; hosted hands-on workshops on Markdown and building with Google Gemini.

Data Engineer (5-person team), USAII Global AI Hackathon 2026

Jun 2026

Wrote the normalization, embedding (gte-small), and batched-insert scripts that loaded 1,008 scraped government schemes into Supabase PostgreSQL/pgvector. Our team was a finalist among 6,081+ participants.

Founder & Editor-in-Chief, Glimpse (student publication)

Jun 2024 – Feb 2025

Founded the publication and ran editorial operations; grew the campus subscriber base to 250+ readers.

Writing

All posts →

Subscribe via RSS →