projects

table of contents

personal

academic

personal

LLM & LLM Serving

Active

During my last quarter at UCSB, I learned a lot about LLM Inference techniques. I decided that the best way to understand how LLMs work and are served is to implement one from scratch. I’m actively traveling, so development is slow, but I’m reading papers and writing code when I can. All code is written by hand (except for some scaffolding). Learning and planning is done in collaboration with Claude.

Links: GitHub

Foundation Model for Prediction Markets

Active

Working on a foundation model for prediction markets. Cannot share much outside of my writings… for obvious reasons.

I do not gamble! However, I have had a lifelong interest in financial markets and derivatives. Hence, I’ve grown to have an interest in prediction markets and their use as a financial tool. Though, most often, they are used as a gambling device.

Linear Encoding of Code Quality in Large Language Models

Apr–Jun 2026 — Course Project for CS 291A, UCSB

We investigate whether code quality metrics are linearly encoded in the activation space of LLMs. Using Contrastive Activation Addition (CAA) and linear probes, we analyze three open-weight models across seven code quality metrics. Linear probes achieve R² ≥ 0.68 for four of the seven metrics, showing strong linear encodings across all three models. Steering vector analysis reveals that the Halstead metrics and Cyclomatic Complexity form a cluster, while Comment Ratio remains largely orthogonal. Steering Comment Ratio and Maintainability Index produces monotonic, controllable changes across most models and steering strengths, with minimal degradation in Pass@1 at moderate α — suggesting that LLMs implicitly learn linear representations of code quality and that steering vectors are a viable training-free mechanism for inference-time control.

Links: Paper

Distributed Data Centers

Jan–Mar 2026 — Python, RAFT, Two-Phase Commit (2PC)

  • Modeled 3 data centers each with 3 servers, responsible for handling financial transactions with strict consistency and atomicity guarantees.
  • Implemented an amalgamation of RAFT consensus and Two-Phase Commit (2PC) for fault-tolerant, atomic distributed transactions.

Links: None… class project, thus, private.

Prediction Market Arbitrage System

Dec 2025–Jan 2026 — Python, asyncio, WebSockets

  • Built a live cross-market arbitrage detection and automated trading system spanning Kalshi and Polymarket prediction markets.
  • Designed core infrastructure: WebSocket-based real-time market data ingestion, event-driven async strategy orchestrator that dispatches trading strategies on incoming signals, and multi-platform order execution via Kalshi and Polymarket APIs.
  • Implemented robust failed-fill handling and concurrency management; system was previously deployed live.

Links: None… private for obvious reasons!

CXXGraph (open-source)

Jul–Sep 2024 — C++, Google Test, CMake

  • Implemented graph algorithms (adjacency and transition matrix powers) for popular header-only C++ library.

Links: GitHub · Presentation

FillerAI

Sep–Dec 2023 — TypeScript, Next.js, GitHub Actions

  • Produced an AI player for a strategy game using Minimax with Alpha-Beta pruning algorithms to make quality moves.
  • Formulated specific mathematically rigorous evaluation function to quantify board states for AI player.

Links: Play! · GitHub

Verde

Dec 2022–Mar 2023 — TypeScript, Firebase, React Native, Expo

  • Crafted a social media app, Verde, with daily environmentally-focused challenges along with photos and user interaction.
  • Contributed with a team of 3 others in an Agile development process using React Native, Expo, and Firebase.
  • Awarded 1st place in UCSB’s Google Developers’ 2023 Solution Challenge.

Links: GitHub

academic

Steering Graph Properties in Large Languages Models

2026 — Graduate Project with Professor Ambuj Singh

this is the project i’m most passionate about, but cannot share much as it’s currently under double-blind review. served as my “project” necessary for MS graduation.

in submission at NeurIPS ‘26. fine-tuned Llama on graph properties. looked at linear probes and steering vectors.

similar to code quality.

AMeLIA: 2-Phase Adversarial Attack on GNNs

2024–2026 — with Professor Ambuj Singh

likely in submission at ICLR ‘27. as i’ve graduated, i’ve stepped away from this project. we developed an adversarial attack framework capable of staying undetected for the first 50% of an adversarial attack while achieving the same degradation as full attacks.

Improving Bounds for Randomly Sampling Graph Colorings

2024–2025 — Undergraduate Senior Thesis with Professor Eric Vigoda

Looked at the classic Sampling for $k$-colorings problem. Developed various variable length and layer coupling approaches with Markov chain Monte Carlo. Lots of linear and quadratic programming!

Links: Poster (formatting is broken)

GraphEval36K: Benchmarking and Improving Large Language Models on Graph Datasets

2024 — with Professor Ambuj Singh

Large language models (LLMs) have achieved remarkable success in natural language processing (NLP), demonstrating significant capabilities in processing and understanding text data. However, recent studies have identified limitations in LLMs’ ability to manipulate, program, and reason about structured data, especially graphs. We introduce GraphEval36K, the first comprehensive graph dataset, comprising 40 graph coding problems and 36,900 test cases to evaluate the ability of LLMs on graph problem-solving. Our dataset is categorized into eight primary and four sub-categories to ensure a thorough evaluation across different types of graphs. We benchmark eight LLMs, finding that private models outperform open-source ones, though the gap is narrowing. We also analyze the performance of LLMs across directed vs undirected graphs, different kinds of graph concepts, and network models. Furthermore, to improve the usability of our evaluation framework, we propose Structured Symbolic Decomposition (SSD), an instruction-based method designed to enhance LLM performance on complex graph tasks. Results show that SSD improves the average passing rate of GPT-4, GPT-4o, Gemini-Pro and Claude-3-Sonnet by 8.38%, 6.78%, 29.28% and 25.28%, respectively.

Links: Paper

Targeted Edge Perturbations on GNNs: Exploring Greedy, Heuristic, and Gradient-Driven Approaches

2023–2024 — ERSP, with Professor Ambuj Singh

We study the Minimum Edge Set Perturbation problem — maximizing edges added to a graph without changing GNN classification accuracy — and introduce the Greedy-Primed Metattack, which achieves comparable accuracy reduction to vanilla Metattack while remaining covert for over 50% of the specified budget.

Links: Research Log · Code · Poster · Proposal