Pittsburgh, Pennsylvania
About
Projects
Blog
Gallery
Ask Anuj
Book

A Researcher & Developer

Hello, I'm Anuj!
I am pursuing a Masters of Computational Data Science at Carnegie Mellon University SCS, researching LLM benchmarks at TEEL Lab, and working as a technical contributor to AspectionLabs.com
Let's Talk

Where my time actually goes

Updated Oct 2026

From Sep 2025 to Sep 2026 I committed 310,866 lines across 997 commits. Most of it went to LLM research (51%), ML systems (14%) and web & apps (14%).

  • LLM research 51%
  • ML systems 14%
  • Coursework 11%
  • Quant finance 3%
  • Web & apps 14%
  • Applied ML 4%
  • Open source & tools 3%

In this window

  1. Private · LLM research 158.2k lines · 395 commits
  2. kvcache-code-rl-optimize (opens in a new tab) 41k lines · 207 commits
  3. Private · Coursework 34.9k lines · 136 commits
  4. This website (opens in a new tab) 17.4k lines · 64 commits
  5. housescope (opens in a new tab) 12.5k lines · 57 commits
  6. Private · Web & apps 12.4k lines · 28 commits
  7. Private · Applied ML 10.5k lines · 39 commits
  8. Private · Open source & tools 8.9k lines · 28 commits
  9. Private · Quant finance 6.3k lines · 7 commits
  10. Private · ML systems 3.1k lines · 11 commits
  11. Everything else 5.8k lines · 25 commits

How I count this

Every line I commit, counted by week, across my own repos, my lab and class repos, and a few projects that never left my laptop. Private work only shows up as its domain. I cap huge commits so one bulk import doesn't pass for a month of work, and I skip lockfiles, generated files and data. It updates on the first of every month.

See the numbers
Lines I committed each month, by domain, with each month's totals. Newest first. Each week counts toward the month it starts in, so the first and last months can be partial.
MonthLLM researchML systemsCourseworkQuant financeWeb & appsApplied MLOpen source & toolsTotal linesCommits
Sep 202664.5k019k04.8k8.5k22397k295
Aug 202619.4k08906021819.7k53
Jul 20265220000005222
Jun 20267.3k0006.2k0013.6k24
May 20269.7k2.8k06.3k00018.8k20
Apr 202624.7k22.9k2.8k2k5.2k2k059.5k296
Mar 202611.2k14.7k6k18720.1k3415.1k57.6k189
Feb 202610.8k3.7k5.7k03.1k3425k28.6k61
Jan 2026000020021
Dec 20257.1k077502.3k0010.2k27
Nov 20251.8k000159042k14
Oct 20251500287041002787410
Sep 20252.4k016000002.5k5
Data (JSON)

RedViz: LLM Red-Teaming Visualization Framework

Streamlit dashboard for probing LLM safety across 8 languages and 9 harm categories with live jailbreak testing and attention analysis.
Read case studyView project

Latest from the blog

Wait, Which PR Got Merged? What Five Closed PRs Taught Me About Open Source

September 27, 2026
Open Source

A New Organism: A Personal Research Report on the State of AI and Its Probable Futures

September 10, 2026
AI/ML

EqRAG: Financial Stock Prediction with Fine-Tuned LLMs

LoRA fine-tuning on 7B math LLMs for Dow 30 stock prediction. Jumped from 12% zero-shot to 51.7% accuracy by training 0.19% of parameters.
Read case studyView project
GAN-Based Synthetic Portrait Generation with Pix2Pix

GAN-Based Synthetic Portrait Generation with Pix2Pix

Three GAN architectures (DCGAN + Pix2Pix) trained to generate anime faces from video frames. 28.5 dB PSNR, 0.91 SSIM.
Read case study
Telecom Customer Churn Prediction: ML Pipeline with Neural Networks

Telecom Customer Churn Prediction: ML Pipeline with Neural Networks

ML pipeline comparing five classifiers on 7,043 telecom customers. Neural network hit 0.83 AUC. Contract type was the strongest churn signal.
Read case study

News Sentiment Analysis with BERT and Web Scraping

NLP pipeline scraping Google News and classifying sentiment with fine-tuned BERT. Includes aspect-level entity extraction.
Read case studyView project
Skin Burn Detection and Classification with YOLOv5

Skin Burn Detection and Classification with YOLOv5

YOLOv5s classifying burn severity into 3 levels on 14,044 medical images. Transfer learning from COCO. Medically validated annotations. Top 5 finalist at the TiE Global Hackathon.
Read case study

Credit Card Fraud Detection

Fraud detection on 284,807 transactions with 0.172% fraud rate. Random Forest achieved 99.96% accuracy. Built without resampling.
Read case studyView project
FIFA Ultimate Team Squad Builder: ML Optimization with Random Forest & MILP

FIFA Ultimate Team Squad Builder: ML Optimization with Random Forest & MILP

Scraped 15,000+ FIFA player cards, predicted goals with Random Forest, and optimized squads with MILP. Acceleration beat shooting stats.
Read case study