Founding Engineer at Starboard · MS Computer Engineering, New York University
Portfolio · LinkedIn · New York, NY
I am a founding engineer at Starboard, where I build multi-agent systems and data pipelines for logistics. My work sits where machine learning meets production engineering: agents that use tools, language models fine-tuned for document understanding, and the infrastructure that keeps them reliable. I hold an MS in Computer Engineering from New York University, where I researched hybrid quantum–classical neural networks at the Center for Quantum Information Physics.
My recent projects span multi-agent reinforcement learning, with a learned interceptor-drone swarm, and medical imaging, with an entry to RSNA's knee MRI challenge on Kaggle; both are described below.
I am also drawn to games, both as engineering problems, as in TicTacPro below, and at the table: I have represented NYU in contract bridge and poker.
Founding Engineer, Starboard · Oct 2025 – Present
Multi-agent systems with tool use that autonomously process shipment tracking, monitor carrier portals, and reconcile data across logistics platforms. Built scrapers and data pipelines with 95%+ accuracy on real-time tracking, and fine-tuned open-source LLMs for document parsing, cutting processing time from 2.5 minutes to under 10 seconds per document.
Research Intern, New York University, Center for Quantum Information Physics · Jan – May 2025
Built a hybrid quantum–classical neural network that estimates six-year lung-cancer risk from a single low-dose CT scan, advised by Prof. Javad Shabani.
Software Development Intern, Tiny Archives · Sep – Dec 2024
Backend archival systems for organizations and individuals in Python, Django, and PostgreSQL.
Software Development Engineer I, Finflux · Nov 2022 – Jul 2023
Built ELMS, an exposure and limit-management engine covering more than 35 million customers, and resolved production issues in the core Loan Management System.
End-to-end encrypted (AES-256-GCM) Web3 messaging with Ethereum and Solana wallet authentication, token-gated communities, and in-chat crypto payments, running on a real-time WebSocket layer with reconnection and delivery confirmation.
Go · React · PostgreSQL · WebSockets
AEGIS trains a swarm of interceptor drones to defend one or more points against incoming attack drones. Coordination is classical and flight is learned: a Hungarian allocator inside the environment assigns each defender an attacker, and one shared PPO policy flies every defender from a 56-dimensional ego-centric observation that does not depend on team size. In simulation the policy wins 98–100% of single-target engagements from 1v1 to 20v20 and holds 100% with zero leakers at 50v40 and when outnumbered 20v40, sizes it never trained at; a single checkpoint defends one to four protected points. I hardened it with adversarial co-training against learned attackers, which exposed and closed a kamikaze-baiting exploit, and built a sim-to-real path to CrazyFlie 2.1 nano-drones that is verified in simulation and software dry runs but has not yet flown.
Python · PyTorch · PPO · NVIDIA Warp · multi-agent reinforcement learning
Code · Paper (PDF) · Competition
An entry to RSNA's Kaggle challenge on knee MRI: 12 findings per study, scored by macro ROC AUC, with radiologist labels for only 58 of 4,407 training studies. The models train on labels that language models extracted from the radiology reports, and an ensemble of 15 EfficientNetV2-S models with a public 2.5D ConvNeXt-T reader scored 0.927 on the private leaderboard. Every change was gated in advance on out-of-fold and gold-study AUC; all 15 submissions kept their order from public to private, and the write-up documents which local gains transferred and which did not.
Python · PyTorch · timm · DICOM
Play the game · Code · Paper (PDF)
TicTacPro is tic-tac-toe with three piece sizes and stacking: a player wins with a line of same-size pieces or a "bullseye" of all three sizes in one cell. An exhaustive alpha-beta solver proves the game is a first-player win from all 27 openings (700 million nodes in 246 seconds on 20 cores), overturning the opposite conclusion that time-limited minimax self-play had converged on. Against the solved game I trained a Dueling Double DQN (792K parameters) on an NVIDIA DGX Spark; a curriculum of rule-based opponents lifted it from 0/20 to 40/40 against the bullseye-fork strategy. The deployed version lets you play against the tree-search AI in the browser.
Python · PyTorch · alpha-beta search · reinforcement learning
A red-teaming entry for a Kaggle competition on tool-using LLM agents (gpt-oss-20b and Gemma). The search finds message chains that make an agent exfiltrate data through otherwise benign tool calls, and reached a public score of 73.88. The write-up also documents why the attack overfit the public guardrail: the hidden private guardrail, which tracks data provenance, blocked it.
Python · llama.cpp · LLM red-teaming
| Project | Summary | Stack |
|---|---|---|
| chromeControl | MCP-compliant Chrome sidebar that reads the current page and answers questions about it, with local processing and domain-based security tiers. | TypeScript, React, MCP |
| Squaris | Three procedurally generated polycube packing puzzles built on one generation kernel, with daily puzzles that are solvable by construction. | React, TypeScript, Three.js |
| Quantum–Classical Neural Network | Six-year lung-cancer risk from a single low-dose CT scan, with model weights stored in qubits through Qiskit. | Python, Qiskit, PyTorch |
| QubitQuery | Retrieval-augmented assistant over lecture notes, using OpenAI embeddings and a FAISS index. | LangChain, FAISS |
| Stock Direction Prediction | Reddit sentiment from fine-tuned FinBERT and Llama models, fed into a Random Forest that forecasts GameStop's price direction. | FinBERT, scikit-learn |
- AEGIS: Learning a Size-Invariant Interceptor Swarm for Drone Defense with a Classical Allocator and One Shared PPO Policy (2026)
- Knee MRI Abnormality Detection with LLM-Read Report Labels: Approaches and Validation Lessons from the RSNA Kaggle Challenge (2026)
- TicTacPro: Exactly Solving a Combinatorial Piece-Placement Game and Stress-Testing Deep Q-Learning Against It (2026)
- Coin-Flip Pricing in a Tournament-Winner Market (2026)
Kaggle: RSNA Knee Abnormality Detection · Kaggle: AI Agent Security, Multi-Step Tool Attacks · HackNYU · Kiro Hackathon (chromeControl) · Reddit Devvit (Squaris) · YQuantum · IMC Prosperity 3 · Hackfinity
New York University, MS in Computer Engineering, 2025
The National Institute of Engineering, Mysuru, BE in Computer Science, 2021