M.S. Robotics Engineering · Widener University
I have never been good at sitting still. I grew up outside, on the water, in the yard, anywhere there was room to move, and even a trip to Disney turned into something other than a vacation. While everyone around me enjoyed the ride, I was quietly pulling it apart in my head, trying to understand what moved it, how it was timed, and why it felt the way it did.
The fun for me was always in that chase. The questions why and how does that work have looped in my head for as long as I can remember, and an easy answer never satisfied me. I needed to find it on my own, and over time that need pointed itself at larger and larger systems.
Never stop asking why!!
The curiosity came naturally. My mom has spent more than thirty years at Merck as a recipe scientist, my dad just as long in automation engineering, and my sister is a practicing OBGYN. I grew up among people who build, measure, and improve things for a living, and a childhood spent on Star Wars did the rest. I wanted to help push the edge of what people can design and build.
Lacrosse is where I learned how to chase that down. From second grade through four years of Division III ball at Widener University, the game shaped me more than any classroom ever did.
Attention to detail.
Lead with passion.
Give everything you have, no matter the scoreboard.
Those habits walked straight into my engineering. Senior year, Dr. Daniel Roozbahani trusted me to lead a ten foot autonomous hexacopter and held the standard punishingly high on purpose. That pressure is the reason I was ready for what came next.
What came next was my master's thesis: a hybrid gasoline and LiPo power system, governed by an intelligent battery management unit, built to stretch the endurance of mobile platforms like robots, marine craft, and drones. I designed and built it on my own, from the switching hardware and firmware up to the scheduling logic that runs it.
Its centerpiece is an algorithm I call Rolling Replenishment that schedules battery swaps proactively instead of reactively. On the bench it recovered 118.6% more usable energy than the best reactive strategy while triggering 58% fewer switching events. I validated a 9 battery prototype on the bench, simulated the full 48 battery configuration in Python, and rebuilt the pack topology from series first to parallel first so the system could run on off the shelf solid state relays with full ground isolation.
In the same window I cofounded CLARO and led every part of its technical build, carrying it from a whiteboard sketch to a shipped iOS app. At its core is a real time engine that flags arbitrage opportunities in under ten seconds across 10 sportsbooks and 22 leagues, wired up with user accounts, push notifications, in app purchases, and deep link bet slips that fill themselves in. It went out to more than 50 beta testers.
“Let the future tell the truth, and evaluate each one according to his work and accomplishments.”— Nikola Tesla
If there is one thing I count on in myself, it is not a single skill. It is that I figure it out. Every project here handed me a problem I could not yet solve, hardware that refused to behave, an algorithm that would not converge, a feature with no clear path. I learned to move toward those walls, take them apart, and keep working until they gave way. An obstacle has never been the reason something of mine went unbuilt.
None of it came easy, and that is the point. No matter how stressful or how hard a project got, my focus, my ability to figure it out, and a driven work ethic are what made it all possible. That same drive is how it all came together in the end, with my master's graduation in 2026, where I finished with a 3.78 GPA.
Hardware or software, physical or financial, the through line holds. Understand the system completely, build it end to end, and prove that it works.
A hardware BMS for extended operation of mobile platforms: 45-channel per-cell monitoring across three Arduino Megas, a solid-state-relay switching network with galvanic isolation, and a custom Rolling Replenishment scheduler. Validated on the bench against six brushless motors.
Led a six-person team building a 10-foot series-hybrid hexacopter with a 4.5 kW onboard generator. Owned the schedule, ~$3K budget, and full system integration, and defined the autonomous battery-rotation concept that became my thesis.
A production iOS app built solo: a sub-10-second detection engine scanning 10 books across 22 leagues, with deep-link auto-fill, auth, push, and in-app purchase. Shipped the full iOS + backend stack end to end to TestFlight beta.
An implied-volatility and Greeks analyzer: a Newton-Raphson IV solver run across live option chains, the full Greeks, SABR volatility-surface modeling, arbitrage detection (put-call parity, calendar, butterfly), and IV-percentile scanners — fronted by a PyQt6 GUI with interactive 3D surface visualization.
A deep-computation March Madness predictor built from scratch: it simulates 88 billion possessions across 10 million tournaments, runs 500,000 MCMC iterations to estimate posterior team-strength distributions, and evolves 1,000 brackets with a genetic algorithm tuned for pool winnings against estimated public pick rates. Finished top 3.5% on ESPN and correctly predicted the champion.
A generative-AI study that has LLMs research a mental illness, then visualize its inner experience — including age-appropriate images built to grow empathy in children.
An end-to-end poker AI: parsed tens of thousands of decisions from Pluribus, engineered poker-specific features, and trained Random Forest + MLP classifiers to imitate its play at ~81% action-prediction accuracy — then wrapped it in a Flask coaching app.
An end-to-end ML framework for forecasting MLB game outcomes. It pulls live data from the official MLB Stats API, builds engineered team- and pitcher-level feature vectors (recent form, run differential, OPS, ERA/WHIP, Pythagorean expectation, pitcher matchup), and trains gradient-boosted trees (XGBoost) against a Random-Forest / Gradient-Boosting / logistic ensemble, scored by a leakage-aware temporal backtester and served through Flask, CLI, and desktop front ends.
Convolutional neural networks for five-class flower classification on a ~4,000-image dataset — 3- and 4-layer architectures with a full preprocessing and augmentation pipeline. The deeper net classified a held-out test set 11/11.
A three-part mobile-robotics series implementing the core autonomy stack — planning, mapping, and localization — in Python and ROS2, demonstrated in Gazebo/RViz simulation: A*/Dijkstra/RRT planning with LQR control, occupancy-grid mapping with graph SLAM, and Monte Carlo Localization.
A MATLAB / Simulink model of an EV drivetrain — DC motor, PI controller, and battery as one system — analyzing power flow through both motoring and regenerative braking.
My master's thesis, Development of a Hybrid Power Generation System Controlled by an Intelligent Battery Management Unit for Extended Operation of Mobile Platforms, defended April 2026 under Dr. Daniel Roozbahani. It set out to push the runtime of mobile and aerial platforms well past the limits of a single battery by intelligently coordinating multiple battery sets alongside an onboard generator.
The core problem is simple to state and hard to solve: a platform that flies or drives on batteries is grounded the moment those batteries drain. Rather than treat the pack as one monolithic source, the BMU treats it as a set of independently switchable banks that can be rotated in and out of service, charged, and monitored while the platform keeps running.
Skywalker III was Widener's senior design capstone: a 10-foot, series-hybrid hexacopter designed to stay aloft far longer than a battery-only drone by carrying its own 4.5 kW generator. As Team Leader of the six-person group, I owned the schedule, the roughly $3K budget, and the full system integration across mechanical, power, and flight-control subsystems.
The defining idea was self-sustaining endurance. Instead of landing to swap or recharge, the airframe carries two battery sets and a generator, rotating them so that one set powers flight while the other charges. That “one set charges, one set flies” concept is what later grew into my master's thesis.
CLARO is a real-time sports arbitrage iOS app I co-founded and built solo, front to back. It is live on TestFlight with 50+ beta users and in App Store review, spanning the full iOS client and its backend services.
The engineering challenge is latency and trust: opportunities exist for seconds, and every one shown to a user has to be real. The detection engine ingests live data, cross-checks it, and surfaces only validated results fast enough to act on.
An implied-volatility and Greeks analyzer that goes from raw option chains to tradable structure: a Newton-Raphson IV solver across every strike and expiration, the full Greeks, SABR volatility-surface modeling, multi-strategy arbitrage detection, and IV-percentile scanners — all behind a desktop GUI with interactive 3D surface visualization.
Live chain data comes from Yahoo Finance (free, no key) with an optional Polygon.io fallback, run through data-quality filters (bid-ask spread, volume, open interest, staleness) and a SQLite cache with rate-limit management.
A deep-computation bracket prediction system for the 2026 NCAA Men's Basketball Tournament, built and run end to end on a single workstation (RTX 5070 Ti, Ryzen, 64 GB DDR5) in 30–45 minutes of real computation per run. It correctly predicted the tournament champion and finished in the 96.5th percentile on ESPN's Tournament Challenge — out of millions of submissions.
This isn't a weighted coin flip across 63 games. The engine simulates actual basketball, possession by possession, and runs six sequential computational epochs to find the optimal bracket.
The final output blends the possession simulation (25%), the MCMC-calibrated Monte Carlo (45%), and the Trapezoid of Excellence pass (30%) into ensemble championship probabilities, then prints the complete bracket — every game, every round, every region — alongside pool-size-specific champion recommendations backed by game-theoretic edge analysis.
A generative-AI project, awarded second place at Widener University's AI Day, built around a deceptively simple question: can a machine with no emotions, no feelings, and no real sight come to understand the inner experience of mental illness well enough to paint it?
The premise was a thought experiment. Take a system that has never felt anxiety, grief, or fear, force it to study the conditions that quietly shape millions of lives, and then ask it to show what it sees. If you could lift someone's internal experience out of their head and set it on a canvas, what would it look like — and could an AI get there?
For each condition — anxiety, depression, PTSD, ADHD, and schizophrenia — the model first ran deep research until it had a complete, grounded understanding of how the illness actually presents. Only then was it asked to generate an image of how it imagined living with that condition from the inside: not a textbook diagram, but an attempt to render the feeling itself. To me, that was the only real way to test whether these models could truly understand and visualize something so human.
The second half turned the research toward education. Some of what drives mental illness is too heavy or too adult to put in front of a child directly — you don't pull a first-grader aside to detail the trauma behind a veteran's flashbacks. But you can say that someone went through something hard, and show an age-appropriate image that helps a child feel the empathetic side: that some people carry invisible struggles. From the other direction, a child who sees themselves in an ADHD image can finally ask for help instead of silently wondering, "what's wrong with me?"
An end-to-end poker AI in two acts. First the model: I parsed tens of thousands of decisions from Pluribus — the superhuman poker bot — engineered poker-specific features, and trained Random Forest and MLP classifiers to imitate its play, reaching ~81% action-prediction accuracy. Then the product: I wrapped that model in a Flask web app that recommends GTO-style actions in real time.
An end-to-end machine-learning framework for forecasting MLB game outcomes and quantifying betting value. It ingests live data from the official MLB Stats API, constructs engineered team- and pitcher-level feature vectors, and trains gradient-boosted trees (XGBoost) alongside a Random-Forest / Gradient-Boosting / logistic-regression ensemble. A leakage-aware temporal backtester — fit strictly on the past and evaluated on a held-out recent window — scores every model on accuracy, precision, recall, and F1, and the whole pipeline is served through a Flask dashboard, a CLI, a desktop GUI, and an automation scheduler over a SQLite tracking store.
A complete, end-to-end ML framework on live MLB data — modeling, leakage-aware backtesting, and a served pipeline. The odds layer is simulated, so it’s an engineering showcase rather than a betting track record.
Convolutional neural networks for five-class flower classification on a dataset of roughly 4,000 images, built in TensorFlow / Keras. I compared 3- and 4-layer architectures with a full preprocessing and augmentation pipeline, then evaluated both with a separate inference script.
Across 15 training epochs the two models reached roughly 74% and 72% validation accuracy, and on a separate held-out test set the four-layer model classified all 11 images correctly (the three-layer, 10 of 11) at high average confidence.
| Image | Actual | Three-layer CNN | Four-layer CNN | ||
|---|---|---|---|---|---|
| Prediction | Conf. | Prediction | Conf. | ||
| daisy_1 | Daisy | Daisy | 100% | Daisy | 100% |
| daisy_2 | Daisy | Daisy | 100% | Daisy | 99.94% |
| dandelion_1 | Dandelion | Dandelion | 56.53% | Dandelion | 51.19% |
| dandelion_2 | Dandelion | Daisy | 50.28% | Dandelion | 43.27% |
| red_sunflower | Sunflower | Sunflower | 99.09% | Sunflower | 97.76% |
| rose_1 | Rose | Rose | 99.62% | Rose | 99.24% |
| rose_2 | Rose | Rose | 98.08% | Rose | 98.51% |
| sunflower_1 | Sunflower | Sunflower | 100% | Sunflower | 99.96% |
| sunflower_2 | Sunflower | Sunflower | 99.98% | Sunflower | 99.02% |
| tulip_1 | Tulip | Tulip | 97.80% | Tulip | 94.47% |
| tulip_2 | Tulip | Tulip | 99.81% | Tulip | 98.45% |
| Totals | 10 / 11 | 91.79% | 11 / 11 | 89.26% | |
11 held-out images. The three-layer net missed one (dandelion read as daisy); the deeper net went 11/11 with marginally lower average confidence.
A three-part mobile-robotics series implementing the core autonomy stack — planning, mapping, and localization — in Python and ROS2, all implemented and demonstrated in Gazebo/RViz simulation. I've pulled the three together into a single report.
A MATLAB / Simulink model of an electric-vehicle drivetrain built to analyze power flow during both motoring and regenerative braking, modeling the DC motor, PI controller, motor controller, and battery as a single system.
As a two-person final project (Team Kappa, RE 403), Mason Reilly and I reconstructed and validated a published EV drive-simulation model, simulating energy flow across a range of speed and torque profiles and tuning the PI loop for closed-loop stability and a sound powertrain control strategy.
My flagship work: the master's thesis and the senior design project it grew out of, told as one continuous story.
My master's thesis: a hybrid gasoline and LiPo power system that keeps drones, robots, EVs, and marine craft running far longer between charges. It closes the energy density gap without giving up the quick response batteries are good at, and I built the whole thing myself, from the battery layout and switching hardware down to the scheduling logic that runs it.
The senior design project that grew into my thesis: a ten-foot, self-charging autonomous hexacopter built to stay in the air longer. I led the team through its design and build, and the endurance problem we kept running into is exactly what my graduate research set out to solve.
View PDF →A competition project that uses AI-generated imagery to show what mental health conditions actually feel like from the inside, covering anxiety, depression, PTSD, ADHD, and schizophrenia. The goal was to build empathy and chip away at stigma, pairing real technical work with a genuinely human focus.
A Bayesian-network model in Python that reads a patient's records and estimates their heart disease risk, returning a probability across graded levels of severity.
An image-classification system I built in Python with convolutional neural networks, trained on a labeled dataset so it can sort new images into the right categories on its own.
View PDF →A deliberate focus on safety culture and ethical responsibility, not coursework for its own sake. Five case studies that dig into how technical flaws, organizational pressure, and ethical lapses combine to cause catastrophe, and what each one asks of the engineers involved.
A close look at what really brought down Challenger in 1986: the failed O-ring seal, a launch decision pushed through under pressure, and the lessons about engineering responsibility that came out of it.
View PDF →A study of the 2003 Columbia re-entry disaster, tracing how foam-strike damage doomed the orbiter and how a culture that kept treating risk as normal let a known hazard slide.
View PDF →A study of the 1988 Piper Alpha explosion, one of the deadliest disasters in oil and gas, and the permit-to-work and safety-system breakdowns that caused it.
View PDF →A study of the 2011 Fukushima Daiichi disaster, looking at how the Tōhoku earthquake and tsunami overwhelmed the plant and how earlier siting and design choices made the outcome worse.
View PDF →A study of the 2023 OceanGate Titan implosion, examining the experimental carbon fiber hull, the independent certification that got skipped, and the ethical failures behind commercial deep-sea tourism.
View PDF →Open to roles across embedded, controls, robotics, software, and quant in the tri-state area or remote.