I’m Noel. I’m a software engineer at Maven AGI, a Boston-based AI startup backed by Dell Technologies Capital, building enterprise AI agents for customer experience. I love working in the startup space because we get to move fast, build cool demos with the newest tech, and then tear them down and rebuild them as real products that customers actually use. Figuring out what’s worth building, what people actually need, and how to make it reliable is the part of the job I find most interesting.
My focus at Maven is developer experience. I build the tooling and systems that help our engineers and customers ship new agents and new capabilities faster. Agent Capabilities is one of those projects. The problem it solves is simple: most teams that deploy a support agent hit the same ceiling fast. The data exists, the systems are connected, but adding new behaviors still means filing a request, waiting on an engineer, and negotiating a sprint slot. Agent Capabilities eliminates that bottleneck — non-technical customers can now build complex agent behaviors themselves, on their own timeline, without touching the engineering queue.
Outside of work, I experiment with agent architectures, proactive behaviors, and what modern AI systems can actually be trusted to do. I’m really interested in AI that’s integrated into products, not just bolted onto them. I also spend a lot of time playing with the newest agent infrastructure — durable execution, agent-native memory, programmatic browsers. Combining those primitives with modern models opens up things that weren’t possible six months ago.
I’m also building Quotient. Tell it what everyone ordered or take a picture of the receipt. It figures out who owes what, splits it however you want, and generates the Venmo requests. The interface is just buttons that do the obvious thing: share a link, text everyone, drop it in the group chat.
Most of what I work on lives in that gap between what AI can do in a demo and what it can do in production. That’s the problem I find most worth working on right now.
8-Year-Old me (left) setting up our robot at the NH FIRST LEGO Robotics competition, my first introduction to engineering.
Projects
Selected Work
All[13]
Featured[5]
Web[4]
Cloud[4]
AI[3]
Embedded[1]
Earlier Work[8]
Featured
Maven AGI – Agent Capabilities
Built the self-serve tooling environment for Maven AGI customers to extend agent behaviors through custom actions and triggers.
TypeScriptReactAPI Design (REST)LLM Orchestration
Agent Capabilities is a self-serve build environment inside Maven AGI's Agent Designer that allows technical teams to quickly create, test, and deploy new agent behaviors. It gives organizations a structured, consistent way to define what agents do and when they do it—using custom actions and triggers. I contributed to building the tooling that makes this possible, enabling customers to extend their AI agents across connected systems without custom development or months-long engineering cycles.
Quotient – AI-Powered Bill Splitting
Built an AI-powered receipt scanning and bill-splitting app with automated Venmo links and SMS reminders.
Quotient is a modern bill-splitting platform designed to eliminate friction in group payments. Users can upload or scan receipts, automatically extract line items using AI-powered OCR, and split expenses intelligently across participants. The system generates clean Venmo deep-links and optional SMS reminders to streamline repayment.
Built with Next.js 14, Supabase (Postgres + pgvector), and Tailwind CSS v4, Quotient leverages Google Gemini for receipt parsing and structured data extraction. The backend is designed for scalability and low-cost operation, with secure user authentication and a clean, mobile-first UI optimized for real-world usage. The goal is to create a frictionless financial coordination layer for everyday group transactions.
Kairos
Built a multi-agent AI planner that breaks down high-level goals into actionable calendar events.
Kairos is an AI-powered weekly planner that intelligently breaks down your high-level goals into actionable calendar events. It leverages a multi-agent orchestration system—comprising fitness, nutrition, and scheduling agents—to generate personalized plans based on your preferences, constraints, and historical data. The front end is built with Next.js and Tailwind CSS for a responsive, glass-like UI, while the FastAPI backend handles agent coordination, RAG-based context retrieval, and integration with Google Calendar. Dockerized and deployed on Google Cloud Run, Kairos uses PostgreSQL with pgvector embeddings to store user profiles and plan references, ensuring scalable, secure, and privacy-focused time management automation.
All Projects
Fall Prevention Education iOS Application
Created an iOS application for the University of Vermont Medical Center to educate fall-prone individuals.
Swift
As part of a collaborative team of four, I contributed to the development of an iOS application for the University of Vermont Medical Center. The objective given to us by the medical professionals, was to educate individuals at risk of falling and empower them to adopt safer practices. Using SwiftUI, we blended design and functional factors to ensure the app's features were tailored to meet the needs of the end-users.
To enhance the app's functionality, we implemented efficient parsing of user data and leveraged locally-stored JSON files for seamless storage and retrieval. Throughout the project, we used an Agile approach, working closely with our team to manage workloads effectively and deliver high-quality results within the allocated semester timeframe.
RNTR - Apartment Rating Website
A review platform for renters in Burlington's captive housing market — built to put information back on the renter's side.
API Design (REST)PostgreSQLDatabase Design
My sophomore year, I went through Burlington's apartment search and it was a nightmare. You'd tour a place, have to decide on the spot, and only after moving in would you find out the utilities were far more expensive than advertised, the landlord was unreachable, and the neighbors were unbearable. There was no way to know any of that before signing a lease.
So I built RNTR, a platform that combined the review model of Yelp with the search and discovery of Zillow, applied specifically to Burlington's rental market. Look up any address and see reviews of the apartment and the landlord, search available rentals, and leave reviews for your current place. The whole idea was that renters deserved to know what they were getting into before they signed anything.
I built the entire thing off the back of one intro web design class, teaching myself everything else from scratch. Authentication flows, relational database design, address standardization, dynamic UI rendering. It was fully functional, tested with friends, and won an award at the UVM CS Fair.
What kept it from gaining traction was the cold start problem. No data meant no users, no users meant no data. Monetization was its own challenge too, since you only apartment hunt once a year and the landlords you're holding accountable have little incentive to support the platform. Eventually let the domain expire.
SG-FECC 2023 Website
Redesigned and implemented an SEO-optimized website, resulting in an estimated 300% increase in page-visits.
As the Senior Media & Communications Chair for the Schlesinger Global Family Enterprise Case Competition, a prestigious global business competition focused on family businesses, I identified the need to enhance our online presence. With participants from 27 countries over the past decade, it was crucial to develop a website that accurately represented the competition's scale and prominence.
Drawing on my web development experience, I quickly mastered WordPress to create a professional and visually appealing website for the current year's competition. This ensured both immediate impact and long-term maintainability, providing a seamless user experience for participants, judges, coordinators, and visitors. I also implemented effective search engine optimization (SEO) techniques, resulting in an estimated 300% increase in website traffic. By optimizing the website's search engine ranking, we significantly expanded the competition's global reach and recognition.
In addition to the website I came up with new ways to promote the competition and engage participants.
Evolutionary Robotics Simulated Robot
Optimized a simulated robot for locomotion in diverse environments using evolutionary algorithms.
Python
Under the guidance of Professor Josh Bongard, I worked on a project to evolve the locomotion capabilities of a simulated robot in various environments. Through evolutionary algorithms, I evolved a neural network to achieve optimal performance. Throughout the semester, I explored different strategies and techniques to continually improve the robot's adaptability and efficiency.
Custom Compiler
Developed a simple compiler enabling translation of high-level programming language into x86 assembly language
Python
For my final project in the course on compiler design and construction, I focused on extending the language's capabilities by implementing simple dataclasses. This project was a culmination of the various topics covered throughout the course, which provided a comprehensive understanding of the compilation process.
Throughout the semester, I explored the representation and analysis of code and how to translate a high-level programming language into Intel x86 assembly language. We learned about essential concepts such as register allocation, static type checking, and handling mutable data. The course also emphasized the significance of garbage collection for efficient memory management in compiled languages.
By extending the language with dataclasses, I not only demonstrated my proficiency in implementing new language features but also showcased my grasp of the entire compilation process. This project solidified my knowledge of code analysis, code generation, and the crucial role of language features in creating powerful and expressive programming languages.
Automated Plant Care System
Using Arduinos, sensors, and pumps connected via Bluetooth, created a device to keep plants watered and fertilized.
C++Python
After failing to keep my plants alive, I used an Arduino Nano and an ESP8266 module, along with sensors for soil moisture, light, and air quality, to automate the process of watering plants and create a more nurturing environment.
By continuously monitoring important factors like soil moisture, light levels, and air quality, my system provided valuable insights into the specific needs of my plants. With the help of a peristaltic pump, the system automatically watered the plants, ensuring they received the appropriate amount of water at the right times.
This project allowed me to explore the integration of hardware components and programming skills, resulting in a practical solution to improve plant care. In the end, I have less dead plants!
Plants vs. Zombies Recreation
Collaborated with a small team to recreate the game Plants vs. Zombies
Python
For this project, I collaborated with an agile team to recreate the immensely popular game, Plants vs. Zombies, using Python and the Python Arcade library. As part of the team, I played a pivotal role in organizing the program flow and backend development. I focused on designing the structure of the game, ensuring seamless transitions between different states and screens. Additionally, I took charge of implementing the spawning mechanics for zombies, creating algorithms that combined randomness with the timing patterns seen in the original game. This resulted in a challenging and engaging gameplay experience.
Another significant contribution I made was developing a dynamic system for level progression. I designed a flexible framework that allowed for the easy addition and modification of levels, accommodating three waves of zombies per level, each increasing in difficulty. This approach added replayability and a sense of progression for players. By working on this project, I enhanced my skills in game development, algorithm design, and collaboration.
Enigma Recreation
Developed a fully-functional version of the WWII German encyphering machine Enigma in Python.
Python
As my first open-ended project during my Introduction to Programming course, I had the ambitious goal of recreating the WWII enciphering machine, Enigma, using Python. Despite being new to programming, I wanted to challenge myself.
Through self-guided learning, I successfully developed a fully functional version of Enigma. Going beyond the original scope, I even added an additional feature that allowed for the translation of enciphered messages into Morse code. This project not only showcased my ability to tackle complex programming tasks but also demonstrated my dedication to expanding my skill set and exploring innovative solutions.
