Helping frontier LLM models write better code and ship faster.
LinkedIn · Portfolio · Google Scholar
I work at the intersection of LLM evaluation, RLHF, and delivery engineering — taking models from promising research to production-ready systems through human-in-the-loop workflows, code benchmarks, and automation that scales. My work has contributed to 25–30%+ gains in model pass rates and coding accuracy for frontier models on real-world SWE tasks. Comfortable across the full stack: Python, TypeScript/JavaScript, React, Node.js, SQL, AWS, and GCP.
- 🔭 Currently building agents and LLM evaluation tooling
- 🎓 MS in Computer Science, Stevens Institute of Technology
- 🌱 Exploring agentic workflows, reward modeling, and from-scratch LLM training
- 📫 Reach me via LinkedIn
| Project | Description | Tech |
|---|---|---|
| llm-as-a-judge-version-2.0 | Full-stack LLM-as-a-Judge platform for evaluating training-dataset quality — rubric extraction, single/pairwise/multi-model debate scoring, ML metrics (Precision/Recall/F1/ROC-AUC), and human review rounds. | TypeScript |
| thriller-short-stories-llm | "Thriller Forge" — a 51M-parameter GPT built and trained from scratch (no pretrained weights) that writes coherent short stories, wrapped in a web app with human-feedback learning. | Python |
| dsawebsite | "Hashmap" — an in-browser DSA learning platform with a Python 3.11 IDE (Pyodide/WebAssembly), AI tutoring, and visual algorithm walkthroughs. 75 problems · 39 lessons · 18 topics. | React · Tailwind |
| split-with-ease | A bill-splitting and personal finance management mobile application. | Mobile |
| cs554-project | "LiveCricketHub" — a live cricket news & score app with ball-by-ball commentary, player stats, live chat rooms, match predictions, and elastic search. | JavaScript |
