Sohan Subudhi

currently

playing

rocket league · pickup basketball

watching

nba playoffs · stand-up specials

making

funny instagram reels with friends

building

gnns over linux traces (ldos lab)

learning

kernel networking, llm code synthesis

Experience

1. Jun 2023 – Aug 2023

   Data Science Intern · VRIZE

   McKinney, TX

   • Wrangled large-scale data to automate training of custom time-series models on a self-service platform (PostgreSQL + S3).
   • Cut model training time by 90% via statistical grid search and AWS parallelization, enabling faster client deployments.
   • Shipped real-time Plotly applications surfacing historical forecasts to non-technical stakeholders.

   vrize.com ↗
2. Jan 2024 – May 2024

   Undergraduate Researcher · Geometry of Space Research Group

   UT Austin

   • Rendered synthetic training images of galaxy mergers from particle data across 10,000+ simulated IllustrisTNG subhalos.
   • Automated a sky-survey API pipeline to curate ~1,000 real-world galaxy merger candidates for manual validation.
   • Integrated Sim2Real domain adaptation with a CNN under Dr. Shyamal Mitra — 40% more accurate merger classification.

   linkedin.com ↗
3. Jan 2025 – Present

   Under Dr. Aditya Akella · NSF Learning-Directed OS Lab, UT Austin

   Undergraduate Researcher

   • Created a training framework for Graph Neural Networks (GNNs) to predict application performance given Linux execution traces.
   • Built an LLM-based code synthesis framework to automatically design Linux kernel TCP congestion algorithms, optimizing low-delay, high-throughput networking for specialized systems.

   ldos.utexas.edu ↗
4. May 2025 – Aug 2025

   Software Engineer Intern · UKG

   Weston, FL

   • Architected the company's first multi-agent AI system (Google ADK), automating cross-functional security remediation through Jira and GitHub.
   • Built high-leverage AI infrastructure with a multi-level Redis caching layer and custom LLM runtime — 12× faster query fulfillment.
   • Maintained reliable real-time technology inventory data via ELT pipelines and a GitHub Actions CI/CD workflow.

   ukg.com ↗

Projects

• Object-Oriented Programming Compiler

  ARM-assembly-targeting compiler with OOP, custom buddy/5th-fit heap manager, and optimizations bringing runtimes within 15% of gcc.

  c++ · arm asm

• 6-Stage Pipelined Processor

  AArch64 subset (misaligned/self-modifying support) with forwarding, dynamic stalls, and speculative branch prediction — 2.1× speedup, CPI 2.8 → 1.3.

  verilog

• Networking Support for Custom OS

  POSIX socket API on a bare-metal kernel — Ethernet/ARP/IPv4/UDP, checksum validation, thread-safe ARP cache, non-blocking I/O.

  c++

• Waste Detection Model

  YOLOv5 satellite-imagery model at 92% mAP for land waste detection. Deployed as a public REST API on Streamlit Cloud.

  python · yolov5

• Donation Prediction Model

  Linear-regression model for the Sankara Eye Foundation predicting donor behavior at 85% MAPE — drove a 12% lift in fundraising efficiency.

  python · sklearn

Education