[{"content":" Attention Visualizer → Attention Visualizer is an interactive tool I built for exploring the self-attention mechanism at the heart of transformer models. It renders tensors and operations as isometric 3D blocks, making it easy to see how shapes flow through each stage of attention — from the initial QKV projections through the softmax and final output projection. You can adjust architecture parameters like sequence length, number of heads, and head dimension in real time and watch the diagram update instantly. Clicking any tensor or operation opens a detail panel with a breakdown of FLOPs, memory transfer, arithmetic intensity, and roofline analysis for A100, H100, and B200 GPUs.\nIf you’re like me and think visually, this is really helpful for understanding the sequence of operations in the major attention variants in use today: MHA, GQA, MQA, and MLA. It’s also helpful for building intuition about how techniques like PagedAttention and FlashAttention speed up inference.\nIn particular, you can answer questions like:\nWhat does the causal attention mask look like for a given set of requests? Check out the mask + softmax op, and be sure to decrease sequence lengths so you can see individual elements visualized. How does Grouped-Query Attention (GQA) help reduce KV cache size compared to Multi-Head Attention (MHA)? The key/value tensors are visibly smaller because there are fewer heads — you can see this directly by adjusting n_kv. How do data parallelism (DP) and tensor parallelism (TP) split up the workload between processors? Increase the sliders to see how the attention heads and requests are divided. Why does MLA use different pipelines for decode and prefill? The crossover analysis shows where the memory-bound to compute-bound transition favors one pipeline over the other. How does FlashAttention avoid materializing the full attention matrix? Toggle it on, and the detail view shows how the computation is tiled across blocks of Q, K, and V. How does PagedAttention organize the KV cache? Toggle it on to see how the cache is broken into fixed-size blocks rather than stored as one contiguous tensor per request. What\u0026rsquo;s the actual compute and memory cost of each operation? Click any op to see a ballpark estimate of the FLOPs, memory transfer, and arithmetic intensity against real GPU specs (A100, H100, B200). ","permalink":"https://matthewbonanni.github.io/posts/2026-05-12-attention-visualizer/","summary":"\u003cdiv style=\"text-align: center; margin-bottom: 1.5rem;\"\u003e\n\u003ca href=\"https://matthewbonanni.github.io/attn-viz/\" style=\"display: inline-flex; align-items: center; gap: 0.5rem; padding: 0.5rem 1.2rem; border: 1px solid var(--border); border-radius: 6px; text-decoration: none; color: var(--primary); font-size: 0.95rem;\"\u003eAttention Visualizer →\u003c/a\u003e\n\u003c/div\u003e\n\u003cp\u003e\u003ca href=\"https://matthewbonanni.github.io/attn-viz/\"\u003eAttention Visualizer\u003c/a\u003e is an interactive tool I built for exploring the self-attention mechanism at the heart of transformer models. It renders tensors and operations as isometric 3D blocks, making it easy to see how shapes flow through each stage of attention — from the initial QKV projections through the softmax and final output projection. You can adjust architecture parameters like sequence length, number of heads, and head dimension in real time and watch the diagram update instantly. Clicking any tensor or operation opens a detail panel with a breakdown of FLOPs, memory transfer, arithmetic intensity, and roofline analysis for A100, H100, and B200 GPUs.\u003c/p\u003e","title":"Attention Visualizer"},{"content":"Overview This was a 6 month design effort by a team of five Northeastern University Mechanical Engineering students: Matthew Bonanni, Max Choate, David Coven, Bryant Grey-Stewart, and Ryan Loehr.\nMy specific contributions encompassed initial requirements analysis, mechanical design of structural components (particularly bearing structures for actuators), and kinematic modeling software.\nAbstract The manipulation of hazardous materials within a glovebox is a promising application of humanoid robotics. NASA R5 \u0026ldquo;Valkyrie\u0026rdquo; is an excellent candidate for this task; however, the constraints of a glovebox limit the volume in which Valkyrie\u0026rsquo;s existing hardware can effectively operate.\nThis project presents assessment of Valkyrie\u0026rsquo;s current forearms and introduces a newly designed forearm optimized for glovebox environments. This design accommodates the Yale OpenHand end effector through compatible mounting. The innovative kinematic configuration increases Valkyrie\u0026rsquo;s glovebox range of motion by more than tenfold while meeting payload and weight specifications.\nGamut Software\n","permalink":"https://matthewbonanni.github.io/posts/2019-05-01-nasa-valkyrie-glovebox-forearms/","summary":"\u003ch2 id=\"overview\"\u003eOverview\u003c/h2\u003e\n\u003cp\u003eThis was a 6 month design effort by a team of five Northeastern University Mechanical Engineering students: Matthew Bonanni, Max Choate, David Coven, Bryant Grey-Stewart, and Ryan Loehr.\u003c/p\u003e\n\u003cp\u003eMy specific contributions encompassed initial requirements analysis, mechanical design of structural components (particularly bearing structures for actuators), and kinematic modeling software.\u003c/p\u003e\n\u003cp\u003e\u003cimg alt=\"Team Photo\" loading=\"lazy\" src=\"/images/posts/valkyrie-glovebox-forearms/team-photo.jpg\"\u003e\u003c/p\u003e\n\u003ch2 id=\"abstract\"\u003eAbstract\u003c/h2\u003e\n\u003cp\u003eThe manipulation of hazardous materials within a glovebox is a promising application of humanoid robotics. NASA R5 \u0026ldquo;Valkyrie\u0026rdquo; is an excellent candidate for this task; however, the constraints of a glovebox limit the volume in which Valkyrie\u0026rsquo;s existing hardware can effectively operate.\u003c/p\u003e","title":"Capstone Project: Forearms for NASA Valkyrie"},{"content":"Capstone Project · GitHub Repo\nOverview This software models and optimizes the accessible volume, or \u0026ldquo;gamut,\u0026rdquo; of robotic arms when constrained by a glovebox.\nI created this tool for our capstone design project involving a glovebox-optimized robotic forearm for NASA\u0026rsquo;s Valkyrie R5 robot. Section 6 of the project\u0026rsquo;s final report contains details on the software\u0026rsquo;s design and implementation.\nRobotic arms are modeled using Denavit-Hartenberg parameters, and represented in MATLAB as rigid body trees. This approach enables computation of transformation matrices between joints at specified angles. The software generates point clouds by iterating through joint position combinations and filtering for collision-free endpoints. Convex hull computation permits design comparisons.\nAn RPR (Roll Pitch Roll) configuration was optimized using brute force methodology. The software adjusts link lengths at set intervals to determine gamut at each iteration. This three-iteration process progressively narrows the parameter space and increases resolution around optimal points for efficient, accurate solutions.\nGamut Visualization of RPR Robotic Forearm\n","permalink":"https://matthewbonanni.github.io/posts/2019-03-01-gamut-modeling/","summary":"\u003cp\u003e\u003ca href=\"/posts/2019-05-01-nasa-valkyrie-glovebox-forearms/\"\u003eCapstone Project\u003c/a\u003e · \u003ca href=\"https://github.com/MatthewBonanni/Glovebox-Forearm-Kinematics\"\u003eGitHub Repo\u003c/a\u003e\u003c/p\u003e\n\u003ch2 id=\"overview\"\u003eOverview\u003c/h2\u003e\n\u003cp\u003eThis software models and optimizes the accessible volume, or \u0026ldquo;gamut,\u0026rdquo; of robotic arms when constrained by a glovebox.\u003c/p\u003e\n\u003cp\u003eI created this tool for our capstone design project involving a glovebox-optimized robotic forearm for NASA\u0026rsquo;s Valkyrie R5 robot. Section 6 of the project\u0026rsquo;s final report contains details on the software\u0026rsquo;s design and implementation.\u003c/p\u003e\n\u003cp\u003eRobotic arms are modeled using Denavit-Hartenberg parameters, and represented in MATLAB as rigid body trees. This approach enables computation of transformation matrices between joints at specified angles. The software generates point clouds by iterating through joint position combinations and filtering for collision-free endpoints. Convex hull computation permits design comparisons.\u003c/p\u003e","title":"Kinematic Modeling of Constrained Robot Arms"},{"content":"Project Site · NASA Site\nOverview As a student researcher in NU\u0026rsquo;s RIVeR (Robotics and Intelligent Vehicles Research) Laboratory, I was initially tasked with organizing a team to design a concept for a new forearm, wrist, and hand for NASA\u0026rsquo;s R5 (Valkyrie) robot.\nThe concept replaces the humanoid design with an active tool change mechanism in the wrist, enabling Valkyrie to attach and detach with multiple \u0026ldquo;hands,\u0026rdquo; each with a specialized purpose.\nWrist My colleague Peter Groen and I designed a concept for the new wrist interface featuring a series of pockets around a cylindrical base, into which pins would be inserted by the wrist to lock it in place.\nWrist Cutaway\nElectrical power and data would interface through the center of the cylinder by means of an orientation-neutral connector.\nThe locking mechanism uses a single motor and features a conical actuator ring which depresses the pins into place while it rotates and slides up on its threads, while remaining robust enough to enable Valkyrie to manipulate heavy objects.\nNew gripper concept, featuring standard wrist interface\n","permalink":"https://matthewbonanni.github.io/posts/2019-01-01-nasa-valkyrie-active-tool-change/","summary":"\u003cp\u003e\u003ca href=\"http://robot.neu.edu/blog/2015/11/19/project-athena-src/\"\u003eProject Site\u003c/a\u003e · \u003ca href=\"https://www.nasa.gov/feature/r5\"\u003eNASA Site\u003c/a\u003e\u003c/p\u003e\n\u003ch2 id=\"overview\"\u003eOverview\u003c/h2\u003e\n\u003cp\u003eAs a student researcher in NU\u0026rsquo;s RIVeR (Robotics and Intelligent Vehicles Research) Laboratory, I was initially tasked with organizing a team to design a concept for a new forearm, wrist, and hand for NASA\u0026rsquo;s R5 (Valkyrie) robot.\u003c/p\u003e\n\u003cp\u003eThe concept replaces the humanoid design with an active tool change mechanism in the wrist, enabling Valkyrie to attach and detach with multiple \u0026ldquo;hands,\u0026rdquo; each with a specialized purpose.\u003c/p\u003e","title":"NASA Valkyrie: Active Tool Change"},{"content":"Full Paper\nOverview Implemented in Simulink, this software models the 2 degree of freedom gantry system shown below. Current is applied to a motor which actuates the cart, and the resulting displacements of the cart and pendulum are measured.\nComparison with experimental data demonstrated this model\u0026rsquo;s accuracy, resulting in a mean cart position error of 0.89cm and pendulum angle error of 0.0217 rad, for a driving frequency of π Hz.\nThe accuracy of the model degenerates, however, with increasing driving frequency due to accumulated frictional losses (bearing friction and drag).\nThis model was developed in satisfaction of the requirements of ME3455: Dynamics and Vibrations.\nComparison of simulated and experimental response to 0.5 Hz driving frequency\n","permalink":"https://matthewbonanni.github.io/posts/2018-09-01-gantry-model/","summary":"\u003cp\u003e\u003ca href=\"/files/me3455_gantry_model_report.pdf\"\u003eFull Paper\u003c/a\u003e\u003c/p\u003e\n\u003ch2 id=\"overview\"\u003eOverview\u003c/h2\u003e\n\u003cp\u003eImplemented in Simulink, this software models the 2 degree of freedom gantry system shown below. Current is applied to a motor which actuates the cart, and the resulting displacements of the cart and pendulum are measured.\u003c/p\u003e\n\u003cp\u003eComparison with experimental data demonstrated this model\u0026rsquo;s accuracy, resulting in a mean cart position error of 0.89cm and pendulum angle error of 0.0217 rad, for a driving frequency of π Hz.\u003c/p\u003e","title":"Dynamic Model of a Gantry System"},{"content":"NU AIAA Project Karman Project Site\nProject Karman is a division of Northeastern University AIAA that is actively developing a rocket designed to breach the Von Karman Line, which defines space.\nOne critical component of the rocket\u0026rsquo;s avionics system is its electronic gyroscope. While the rocket is in flight, the gyroscope collects data to ensure it is oriented properly. If the rocket veers off course, the on-board avionics system will prevent the second stage from firing, thereby minimizing the rocket\u0026rsquo;s lateral travel.\nMy first task on this team was to develop a test fixture ensuring the gyroscope will work properly while the rocket is rapidly rotating in flight.\nDesign The fixture was designed to mimic a three-axis gimbal. A motor turns a central axle to which the gyroscope is mounted, rotating it at approximately 10 Hz. This unit is allowed to roll and pitch freely by the gimbals.\nAs the avionics are spun, the central unit is tilted by hand. Encoders measure the roll and pitch angles, and this data is compared to the gyroscope\u0026rsquo;s measurements to ensure it is functioning properly.\n","permalink":"https://matthewbonanni.github.io/posts/2018-06-01-gyro-test-fixture/","summary":"\u003ch2 id=\"nu-aiaa-project-karman\"\u003eNU AIAA Project Karman\u003c/h2\u003e\n\u003cp\u003e\u003ca href=\"https://web.northeastern.edu/aiaa/project-karman/\"\u003eProject Site\u003c/a\u003e\u003c/p\u003e\n\u003cp\u003eProject Karman is a division of Northeastern University AIAA that is actively developing a rocket designed to breach the Von Karman Line, which defines space.\u003c/p\u003e\n\u003cp\u003eOne critical component of the rocket\u0026rsquo;s avionics system is its electronic gyroscope. While the rocket is in flight, the gyroscope collects data to ensure it is oriented properly. If the rocket veers off course, the on-board avionics system will prevent the second stage from firing, thereby minimizing the rocket\u0026rsquo;s lateral travel.\u003c/p\u003e","title":"Gyroscope Test Fixture"},{"content":"Project Site\nOverview In Fall 2017, I began working on Paradigm Hyperloop, an international team of students from Northeastern University and Memorial University of Newfoundland \u0026amp; Labrador.\nThe team previously achieved second place in Elon Musk\u0026rsquo;s Hyperloop Pod Competition II and was the only North American finalist. We were developing a new pod for the next competition.\nMy initial responsibilities included redesigning the pod\u0026rsquo;s suspension system with focus on design for manufacturing (DFM) and simplified assembly.\nDesign The Paradigm Hyperloop Pod relies on air skate levitation. Air is forced through the red nylon airbags, drastically reducing friction between the pod and the track. When no air is supplied, the pod rests on its wheels.\nAir bearings are the core technology we have developed for Competition II. While air bearings are already utilized in low speed industrial applications, they\u0026rsquo;ve never been explored in high-speed contexts. Low-friction levitation is critical to the Hyperloop concept and our air bearings reduce the force necessary to propel our pod by 80%.\nThe pod uses four air skate assemblies, each equipped with independent suspension featuring a parallel four-bar linkage with centralized damped spring.\nImproved Suspension\n","permalink":"https://matthewbonanni.github.io/posts/2018-01-01-paradigm-hyperloop/","summary":"\u003cp\u003e\u003ca href=\"https://paradigmhyperloop.com/\"\u003eProject Site\u003c/a\u003e\u003c/p\u003e\n\u003ch2 id=\"overview\"\u003eOverview\u003c/h2\u003e\n\u003cp\u003eIn Fall 2017, I began working on Paradigm Hyperloop, an international team of students from Northeastern University and Memorial University of Newfoundland \u0026amp; Labrador.\u003c/p\u003e\n\u003cp\u003eThe team previously achieved second place in Elon Musk\u0026rsquo;s Hyperloop Pod Competition II and was the only North American finalist. We were developing a new pod for the next competition.\u003c/p\u003e\n\u003cp\u003eMy initial responsibilities included redesigning the pod\u0026rsquo;s suspension system with focus on design for manufacturing (DFM) and simplified assembly.\u003c/p\u003e","title":"Paradigm Hyperloop"},{"content":"GitHub Repo\nOverview I created software for addressing Pareto optimization challenges using MATLAB, complete with a sample mathematical model.\nPareto optimization problems are those in which there is more than one objective for optimization. In complex scenarios, enhancing one objective often requires compromise on others, creating an optimal surface rather than a single solution point. Any point on the surface is an optimal solution to the problem. When relative weights are established, each point receives a weighted score to identify the singular optimal solution.\nThe sample model features four independent parameters and three optimization objectives.\nBrute Force This method examines all parameter combinations to determine output objectives. Normalization scales all values between 0 and 1 by dividing the objective values at each point by this maximum. The weighted score of each point is represented by its color. 3 of the parameters are represented as physical axes, while the fourth is varied over time.\nVisualization of Pareto Optimization\nGradient Descent This more efficient approach uses a weighted scoring system. The algorithm starts at a given seed point and initially determines Utopia and Nadir points through independent objective optimization.\nAt this point, it runs the mathematical model over all surrounding points by incrementing and decrementing each parameter. The cursor moves to the point with the highest score.\nThe method repeats until reaching peak performance, with multiple random seed points tested in an effort to escape false peaks (local maxima).\nGradient Descent Visualization\nGradient Descent - Top View\nResults Both techniques were successfully demonstrated. The brute force approach proved slower but reliable, while gradient descent operated faster but showed difficulty in converging on a singular solution.\nThis was my first software development project of this complexity and scale, and it significantly enhanced my expertise in mathematical modeling and optimization methodologies, skills later applied to mechanical design challenges.\n","permalink":"https://matthewbonanni.github.io/posts/2017-06-01-pareto-optimization/","summary":"\u003cp\u003e\u003ca href=\"https://github.com/MatthewBonanni/Optimization-Model\"\u003eGitHub Repo\u003c/a\u003e\u003c/p\u003e\n\u003ch2 id=\"overview\"\u003eOverview\u003c/h2\u003e\n\u003cp\u003eI created software for addressing Pareto optimization challenges using MATLAB, complete with a sample mathematical model.\u003c/p\u003e\n\u003cp\u003ePareto optimization problems are those in which there is more than one objective for optimization. In complex scenarios, enhancing one objective often requires compromise on others, creating an optimal surface rather than a single solution point. Any point on the surface is an optimal solution to the problem. When relative weights are established, each point receives a weighted score to identify the singular optimal solution.\u003c/p\u003e","title":"Pareto Optimization Software"},{"content":"Implemented in MATLAB, this software computes projectile motion trajectories using both viscous and Newtonian models of air resistance. Equations of motion are solved numerically using the Runge-Kutta method.\nThe project originated from coursework for PHYS 3601: Classical Mechanics and evolved to include a user-friendly graphic interface and the ability to process many sets of input parameters through .csv file import.\nGitHub Repo\n","permalink":"https://matthewbonanni.github.io/posts/2017-01-01-projectile-motion-calculator/","summary":"\u003cp\u003eImplemented in MATLAB, this software computes projectile motion trajectories using both viscous and Newtonian models of air resistance. Equations of motion are solved numerically using the Runge-Kutta method.\u003c/p\u003e\n\u003cp\u003eThe project originated from coursework for PHYS 3601: Classical Mechanics and evolved to include a user-friendly graphic interface and the ability to process many sets of input parameters through .csv file import.\u003c/p\u003e\n\u003cp\u003e\u003ca href=\"https://github.com/MatthewBonanni/Projectile-Motion-Calculator\"\u003eGitHub Repo\u003c/a\u003e\u003c/p\u003e","title":"Projectile Motion Calculator"},{"content":"View Paper\nOverview The Door Motion Generator is a device designed to generate electricity from the motion of opening doors. It was created by five freshman students as a final project for GE 1110. The corresponding paper for this project went on to win Best Undergraduate Research Paper at the ASEE Northeast Regional Conference in 2016.\nThe project aimed to demonstrate everyday energy expenditure and served as mechanical design practice early in undergraduate studies.\nDesign The device mounts to a door frame, functioning like a lawnmower pull-start. As the door opens, a cable pulls and rotates a motor. A centrifugal clutch enables the motor to drive when pulled and retracts the cable easily via coil spring as the door closes.\nI developed all of its mechanisms, including the pulley system and clutch, sourcing a coil spring from a chainsaw starter, and designing housing and mounting hardware. Due to time and budget constraints, only one prototype was built. Design iteration occurred purely in CAD and was largely focused on reducing the overall size.\nExploded View\nConclusions The prototype served as an effective proof of concept despite not generating meaningful power. The centrifugal clutch could be replaced with a full-wave rectifier to simplify the system with fewer moving parts.\nFull Wave Rectifier\nSome doors may see upwards of 1500 opening cycles per day, supporting the concept\u0026rsquo;s potential with improved design.\nThe Door Motion Generator team with our advisor, Dr. Maheswaran, at ASEE NE 2016. (Not pictured: Spencer Pozder)\n","permalink":"https://matthewbonanni.github.io/posts/2016-05-01-door-motion-generator/","summary":"\u003cp\u003e\u003ca href=\"/files/door_motion_generator.pdf\"\u003eView Paper\u003c/a\u003e\u003c/p\u003e\n\u003ch2 id=\"overview\"\u003eOverview\u003c/h2\u003e\n\u003cp\u003eThe Door Motion Generator is a device designed to generate electricity from the motion of opening doors. It was created by five freshman students as a final project for GE 1110. The corresponding paper for this project went on to win Best Undergraduate Research Paper at the ASEE Northeast Regional Conference in 2016.\u003c/p\u003e\n\u003cp\u003eThe project aimed to demonstrate everyday energy expenditure and served as mechanical design practice early in undergraduate studies.\u003c/p\u003e","title":"Door Motion Generator"},{"content":" Stanford University PhD, Mechanical Engineering · 2019 – 2025\nMS, Mechanical Engineering · 2019 – 2021\nDoctoral Advisor: Matthias Ihme · FX Lab\nThesis: Multiscale modeling of high-speed propulsion systems\nFifth Year Spring 2024:\nME 375: Wildfire Science (TA) Winter 2024:\nME 335A: Finite Element Analysis Fall 2023:\nAA 228: Decision Making Under Uncertainty Fourth Year Spring 2023:\nME 375: Wildfire Science (TA) ME 451C: Low-Order Modeling for Turbulent Flows Winter 2023:\nAA 283: Aircraft and Rocket Propulsion Fall 2022:\nME 451B: Flow Instabilities Third Year Spring 2022:\nME 339: Parallel Computing Winter 2022:\nME 451A: Multiphase Flows Fall 2021:\nCS 230: Deep Learning Second Year Spring 2021:\nME 361: Turbulence ME 471: Turbulent Combustion Winter 2021:\nAA 214: Numerical Methods for Compressible Flows CS 205L: Continuous Mathematical Methods with Emphasis on Machine Learning Fall 2020:\nAA 210A: Compressible Flow First Year Spring 2020:\nME 300C: Numerical Methods ME 372: Combustion Applications Winter 2020:\nME 300B: Partial Differential Equations ME 351B: Viscous Fluid Mechanics ME 371: Combustion Fundamentals Fall 2019:\nME 300A: Linear Algebra ME 351A: Inviscid Fluid Mechanics ME 362A: Physical Gas Dynamics Northeastern University BS, Mechanical Engineering · Honors Program · 2015 – 2019\nMinor, Physics\nSumma Cum Laude · GPA: 3.99 / 4.00\nFourth Year Spring 2019:\nME 4508: Mechanical Engineering Computation and Design (FEA) ME 4565: Introduction to Computational Fluid Dynamics ME 4505: Measurement and Analysis with Thermal Science Application ENGW 3302: Advanced Writing in the Technical Professions Fall 2018:\nME 4555: Systems Analysis and Control ME 4570: Thermal Systems Analysis and Design (Heat Transfer) ME 7210: Elasticity and Plasticity (Graduate Level) MEIE 4702: Capstone II MEIE 3000: Professional Issues in Engineering Third Year Summer II 2018:\nME 3475: Fluid Mechanics ME 4550: Mechanical Engineering Design ME 4701: Capstone I Spring/Summer I 2018:\nSix-month co-op at SpaceX Fall 2017:\nME 3455: Dynamics and Vibrations ME 2340: Introduction to Materials Science PHYS 2371: Electronics CS 2500: Fundamentals of Computer Science Second Year Spring/Summer I 2017:\nSix-month co-op at iRobot Fall 2016:\nME 2380: Thermodynamics ME 2355: Mechanics of Materials PHYS 3601: Classical Dynamics MATH 2341: Differential Equations and Linear Algebra First Year Summer I/II 2016:\nTen-week internship at GE Global Research Spring 2016:\nME 2350: Engineering Mechanics and Design (Statics) MATH 2321: Calculus III PHYS 2303: Modern Physics HONR 1208: Honors Seminar: What Makes Music Work Fall 2015:\nCHEM 1151: Chemistry for Engineers MATH 1342: Calculus II GE 1110: Engineering Design GE 1111: Problem Solving and Computation GE 1000: Introduction to the Study of Engineering HONR 1101: Enhancing Honors Pre-College Coursework Hudson Valley Community College PHY 150: Physics I PHY 151: Physics II Spanish V SUNY Albany Spanish IV AP Courses Calculus AB Microeconomics Literature US History European History ","permalink":"https://matthewbonanni.github.io/education/","summary":"\u003cdiv class=\"experience-entry\"\u003e\n\u003cdiv class=\"experience-logo\"\u003e\n\u003cimg src=\"/images/education/stanford-seal.png\" alt=\"Stanford University\"\u003e\n\u003c/div\u003e\n\u003cdiv class=\"experience-details\"\u003e\n\u003ch3 id=\"stanford-university\"\u003eStanford University\u003c/h3\u003e\n\u003cp\u003e\u003cstrong\u003ePhD, Mechanical Engineering\u003c/strong\u003e · 2019 – 2025\u003cbr\u003e\n\u003cstrong\u003eMS, Mechanical Engineering\u003c/strong\u003e · 2019 – 2021\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eDoctoral Advisor:\u003c/em\u003e Matthias Ihme · \u003ca href=\"https://fxlab.stanford.edu\"\u003eFX Lab\u003c/a\u003e\u003cbr\u003e\n\u003cem\u003eThesis:\u003c/em\u003e \u003ca href=\"https://purl.stanford.edu/cg360dx4750\"\u003eMultiscale modeling of high-speed propulsion systems\u003c/a\u003e\u003c/p\u003e\n\u003ch4 class=\"coursework-year\"\u003eFifth Year\u003c/h4\u003e\n\u003cp class=\"coursework-term\"\u003eSpring 2024:\u003c/p\u003e\n\u003cul class=\"coursework-courses\"\u003e\n\u003cli\u003e\u003cstrong\u003eME 375:\u003c/strong\u003e Wildfire Science (TA)\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp class=\"coursework-term\"\u003eWinter 2024:\u003c/p\u003e\n\u003cul class=\"coursework-courses\"\u003e\n\u003cli\u003e\u003cstrong\u003eME 335A:\u003c/strong\u003e Finite Element Analysis\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp class=\"coursework-term\"\u003eFall 2023:\u003c/p\u003e\n\u003cul class=\"coursework-courses\"\u003e\n\u003cli\u003e\u003cstrong\u003eAA 228:\u003c/strong\u003e Decision Making Under Uncertainty\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch4 class=\"coursework-year\"\u003eFourth Year\u003c/h4\u003e\n\u003cp class=\"coursework-term\"\u003eSpring 2023:\u003c/p\u003e\n\u003cul class=\"coursework-courses\"\u003e\n\u003cli\u003e\u003cstrong\u003eME 375:\u003c/strong\u003e Wildfire Science (TA)\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eME 451C:\u003c/strong\u003e Low-Order Modeling for Turbulent Flows\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp class=\"coursework-term\"\u003eWinter 2023:\u003c/p\u003e\n\u003cul class=\"coursework-courses\"\u003e\n\u003cli\u003e\u003cstrong\u003eAA 283:\u003c/strong\u003e Aircraft and Rocket Propulsion\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp class=\"coursework-term\"\u003eFall 2022:\u003c/p\u003e","title":"Education"},{"content":" Red Hat Boston, MA\nMachine Learning Engineer\nJuly 2025 – Present\nContributing to and maintaining vLLM, the premier open-source LLM inference engine.\nNASA Ames Research Center, Mountain View, CA\nVisiting NASA Fellow (CBRE)\nJune 2024 – Sept. 2024\nDeveloped adaptive modeling framework for radiative heat transfer in hypersonic reentry simulations.\nGlenn Research Center, Cleveland, OH\nVisiting NASA Fellow (CBRE)\nJune 2023 – Sept. 2023\nCreated fidelity-adaptive scramjet combustion models achieving 2x computational speedup while maintaining accuracy; provided guidance on combustion model implementation in NASA CFD software.\nLangley Research Center, Hampton, VA\nVisiting NASA Fellow (CBRE)\nJune 2022 – Sept. 2022\nExecuted large-eddy simulations of liquid-fueled scramjet combustor; created analysis methods for cavity residence time statistics; built low-order combustion models.\nSpaceX Hawthorne, CA\nAssociate Engineer, Post Grad — Falcon Structures Engineering\nMay 2019 – Aug. 2019\nDesigned structural hardware for Falcon 9 landing legs; analyzed Falcon Heavy core structure; created software tools for design automation; conducted pressurized testing on Stage 2 hardware.\nVehicle Engineering Co-op — Falcon Integration and Test\nJan. 2018 – June 2018\nImplemented design improvements for Falcon 9 Stage 1 reusability; created Dragon 2 weld inspection tool reducing setup time by 74%; authored corrosion control documentation; designed test fixtures for foam and coating systems.\niRobot Bedford, MA\nMechanical Engineering Co-op — Wet Floor Care Robotics\nJan. 2017 – June 2017\nLed design of Braava Jet m6 production components; developed optimization software for multi-link mechanisms; refined designs balancing strength, stability, and manufacturing constraints.\nGE Global Research Niskayuna, NY\nIntern — Advanced Communications Systems \u0026amp; Distributed Intelligent Systems Labs\nJune 2016 – Sept. 2016\nDeveloped image processing software reducing photogrammetry labor costs by 85%; directed mechanical design of turbine repair robot, enhancing stability and modularity.\n","permalink":"https://matthewbonanni.github.io/experience/","summary":"\u003cdiv class=\"experience-entry\"\u003e\n\u003cdiv class=\"experience-logo\"\u003e\n\u003cimg src=\"/images/experience/redhat-logo.svg\" alt=\"Red Hat\"\u003e\n\u003c/div\u003e\n\u003cdiv class=\"experience-details\"\u003e\n\u003ch3 id=\"red-hat\"\u003eRed Hat\u003c/h3\u003e\n\u003cp\u003eBoston, MA\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMachine Learning Engineer\u003c/strong\u003e\u003cbr\u003e\n\u003cstrong\u003eJuly 2025 – Present\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eContributing to and maintaining \u003ca href=\"https://github.com/vllm-project/vllm\"\u003evLLM\u003c/a\u003e, the premier open-source LLM inference engine.\u003c/p\u003e\n\u003c/div\u003e\n\u003c/div\u003e\n\u003chr\u003e\n\u003cdiv class=\"experience-entry\"\u003e\n\u003cdiv class=\"experience-logo\"\u003e\n\u003cimg src=\"/images/experience/nasa-logo.png\" alt=\"NASA\"\u003e\n\u003c/div\u003e\n\u003cdiv class=\"experience-details\"\u003e\n\u003ch3 id=\"nasa\"\u003eNASA\u003c/h3\u003e\n\u003cp\u003eAmes Research Center, Mountain View, CA\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eVisiting NASA Fellow (CBRE)\u003c/strong\u003e\u003cbr\u003e\n\u003cstrong\u003eJune 2024 – Sept. 2024\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eDeveloped adaptive modeling framework for radiative heat transfer in hypersonic reentry simulations.\u003c/p\u003e\n\u003chr\u003e\n\u003cp\u003eGlenn Research Center, Cleveland, OH\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eVisiting NASA Fellow (CBRE)\u003c/strong\u003e\u003cbr\u003e\n\u003cstrong\u003eJune 2023 – Sept. 2023\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eCreated fidelity-adaptive scramjet combustion models achieving 2x computational speedup while maintaining accuracy; provided guidance on combustion model implementation in NASA CFD software.\u003c/p\u003e","title":"Experience"},{"content":" Publications M. Bonanni, A. Norris, M. Ihme, \u0026ldquo;Analysis of residence time distribution in a cavity-stabilized scramjet combustor,\u0026rdquo; Proceedings of the Combustion Institute, 2024. DOI\nM. Bonanni, D. Brouzet, G. Vignat, M. Ihme, \u0026ldquo;Coupling of detonation structure and upstream inhomogeneities in a rotating detonation engine,\u0026rdquo; Proceedings of the Combustion Institute, 2024. DOI\nR. Chen, M. Bonanni, M. Ihme, \u0026ldquo;Low-Order Model for Generating Supersonic Cavity Combustion Ignition Probability Maps,\u0026rdquo; AIAA Aviation 2024 Forum, 2024. DOI\nG. Vignat, D. Brouzet, M. Bonanni, M. Ihme, \u0026ldquo;Analysis of weak secondary waves in a rotating detonation engine,\u0026rdquo; Combustion and Flame, 2024. DOI\nJ. Burge, M. Bonanni, L. Hu, M. Ihme, \u0026ldquo;Recurrent Convolutional Deep Neural Networks for Modeling Time-Resolved Wildfire Spread Behavior,\u0026rdquo; Fire Technology, 2023. DOI\nM. Bonanni, A. Norris, M. Ihme, \u0026ldquo;Adaptive Modeling of Supersonic Combustion in a Cavity-Stabilized Scramjet,\u0026rdquo; AIAA Scitech 2023 Forum, 2023. DOI\nM. Bonanni, M. Ihme, \u0026ldquo;Interaction of preferential evaporation and low-temperature chemistry in multicomponent counterflow spray flames,\u0026rdquo; Proceedings of the Combustion Institute, 2023. DOI\nM. Bonanni et al., \u0026ldquo;Toward Numerical Investigation of Ignition and Combustion Transition in a Subscale LOX/Methane Rocket Combustor,\u0026rdquo; AIAA Scitech 2021 Forum, 2021. DOI\nJ. Burge, M. Bonanni, M. Ihme, L. Hu, \u0026ldquo;Convolutional LSTM Neural Networks for Modeling Wildland Fire Dynamics,\u0026rdquo; arXiv, 2020. arXiv:2012.06679\nConference Presentations B. Andersen, J. Z. Ho, G. Vignat, M. Bonanni, D. Brouzet, M. Ihme, \u0026ldquo;Large Eddy Simulation of Turbulent Partially Premixed NH3-Air Flames,\u0026rdquo; Hydrogen Innovation and Technology Conference, 2025.\nB. Andersen, J. Z. Ho, G. Vignat, M. Bonanni, H. Tang, G. Magnotti, M. Ihme, \u0026ldquo;Large eddy simulation of piloted turbulent partially premixed NH3/H2/N2-air jet flames,\u0026rdquo; National Combustion Meeting, 2025.\nM. Bonanni, N. Perakis, A. Norris, M. Ihme, \u0026ldquo;Prediction of heat release and wall heat transfer in a scramjet combustor via Pareto-efficient combustion modeling,\u0026rdquo; International Conference on Numerical Combustion, 2024.\nM. Bonanni, D. Brouzet, G. Vignat, M. Ihme, \u0026ldquo;Local Dynamic Combustion Model Adaptation for Large-Eddy Simulation of Scramjets at Reduced Cost,\u0026rdquo; Int. Colloquium on the Dynamics of Explosions and Reactive Systems, 2023.\nM. Bonanni, D. Brouzet, G. Vignat, M. Ihme, \u0026ldquo;Examining Structural Inhomogeneities of Detonations in a Rotating Detonation Rocket Engine,\u0026rdquo; Int. Colloquium on the Dynamics of Explosions and Reactive Systems, 2023.\nG. Vignat, D. Brouzet, M. Bonanni, M. Ihme, \u0026ldquo;Interaction Between Primary and Secondary Waves in a Rotating Detonation Rocket Engine,\u0026rdquo; Int. Colloquium on the Dynamics of Explosions and Reactive Systems, 2023.\nM. Bonanni, N. Perakis, A. Norris, M. Ihme, \u0026ldquo;Large-eddy simulations of a cavity-stabilized scramjet with Pareto-efficient combustion modeling,\u0026rdquo; TFSA, 2023.\nM. Bonanni, D. Brouzet, G. Vignat, J. J. Hansen, M. Ihme, \u0026ldquo;Macroscopic effects of numerical schemes in an RDE,\u0026rdquo; Model Validation for Propulsion Workshop, AIAA SciTech 2023 Forum, 2023.\nM. Bonanni, \u0026ldquo;Eulerian-Lagrangian LES analysis of residence time in a scramjet cavity combustor,\u0026rdquo; Bulletin of the American Physical Society (APS Division of Fluid Dynamics), 2022. Link\nM. Bonanni, \u0026ldquo;Toward Fidelity-Adaptive Simulation of a Cavity-Stabilized Scramjet Combustor,\u0026rdquo; Meeting of the Western States Section of the Combustion Institute, 2022.\nG. Vignat, D. Brouzet, M. Bonanni, M. Ihme, \u0026ldquo;Large eddy simulation of a rotating detonation rocket engine,\u0026rdquo; Meeting of the Western States Section of the Combustion Institute, 2022.\nM. Bonanni, \u0026ldquo;One-Dimensional Simulations of Multicomponent Counterflow Spray Flames,\u0026rdquo; Bulletin of the American Physical Society (APS Division of Fluid Dynamics), 2021. Link\nM. Bonanni, M. Börner, J. Hardi, M. Ihme, \u0026ldquo;Turbulent Spray Combustion with Flash Evaporation during Rocket Combustor Startup,\u0026rdquo; EUROMECH Colloquium 621: Transport and fluxes in dispersed turbulent flows, 2021.\nM. Bonanni, \u0026ldquo;Ensemble Predictions of Wildfire Spread Through TPU-Compatible TensorFlow Acceleration,\u0026rdquo; Bulletin of the American Physical Society (APS Division of Fluid Dynamics), 2020. Link\nGoogle Scholar ","permalink":"https://matthewbonanni.github.io/publications/","summary":"\u003cdiv style=\"text-align: center; margin-bottom: 2rem;\"\u003e\n\u003ciframe width=\"800\" height=\"450\" src=\"https://www.youtube.com/embed/TJsjlOik0QU\" frameborder=\"0\" allowfullscreen style=\"max-width: 100%;\"\u003e\u003c/iframe\u003e\n\u003c/div\u003e\n\u003ch2 id=\"publications\"\u003ePublications\u003c/h2\u003e\n\u003col\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eM. Bonanni\u003c/strong\u003e, A. Norris, M. Ihme, \u0026ldquo;Analysis of residence time distribution in a cavity-stabilized scramjet combustor,\u0026rdquo; \u003cem\u003eProceedings of the Combustion Institute\u003c/em\u003e, 2024. \u003ca href=\"https://doi.org/10.1016/j.proci.2024.105690\"\u003eDOI\u003c/a\u003e\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003e\u003cstrong\u003eM. Bonanni\u003c/strong\u003e, D. Brouzet, G. Vignat, M. Ihme, \u0026ldquo;Coupling of detonation structure and upstream inhomogeneities in a rotating detonation engine,\u0026rdquo; \u003cem\u003eProceedings of the Combustion Institute\u003c/em\u003e, 2024. \u003ca href=\"https://doi.org/10.1016/j.proci.2024.105576\"\u003eDOI\u003c/a\u003e\u003c/p\u003e\n\u003c/li\u003e\n\u003cli\u003e\n\u003cp\u003eR. Chen, \u003cstrong\u003eM. Bonanni\u003c/strong\u003e, M. Ihme, \u0026ldquo;Low-Order Model for Generating Supersonic Cavity Combustion Ignition Probability Maps,\u0026rdquo; \u003cem\u003eAIAA Aviation 2024 Forum\u003c/em\u003e, 2024. \u003ca href=\"https://doi.org/10.2514/6.2024-3809\"\u003eDOI\u003c/a\u003e\u003c/p\u003e","title":"Publications"}]