Sparsh Kudrimoti — Projects

Explore by discipline

Computer Science

Machine learning & software

Splor, ultrasound deep learning, NeuroPrep, and a peer-reviewed clustering study.

Mechanical Engineering

Medical devices, robotics & CAD

MediFlex Bell, a sensor mount for a robot dog, and hardware internships — CAD, FEA, and prototyping.

Business & Finance

Product, growth & markets

Product management at Capital One, equity research, and angel investing at Demotu.

Projects & experience

Computer Science

Machine learning, data science, and software — models, apps, and a peer-reviewed publication.

Venture

Splor: a social travel app for iOS

A full-stack iOS app for planning trips with friends, discovering places, and staying motivated through streaks and friendly competition.

Role: Founding Developer
When: Oct 2025 – May 2026
Stack: Swift, SwiftUI, Supabase, Gemini

What I built

A Swift/SwiftUI app backed by a Supabase database and storage with real-time sync
Apple Sign-In via OAuth
An AI concierge search feature, powered by Gemini, for personalized recommendations

Social features

Gamification with streaks and leaderboards
Activity feeds and trip collaboration tools
Pair-programmed with Claude Code to ship faster

Physiology of Wearable Robotics Lab

Automating ultrasound muscle analysis with machine learning

Measuring muscle in ultrasound images used to mean tracing every frame by hand. I built a MATLAB deep-learning tool that finds the top and bottom muscle borders (aponeuroses) automatically — and that other labs can quickly retrain for their own studies.

Role: Machine Learning Engineer
Lab: PoWeR Lab, Georgia Tech (Prof. Sawicki)
When: Aug 2020 – May 2024
Stack: MATLAB, Deep Learning Toolbox, ImageJ

99.5%Global pixel accuracy (top border)

0.99Weighted IoU, top border model

490Hand-labeled training images, 7 participants

~800 hrsManual labeling time saved

Pipeline

Collect raw ultrasound→ Label borders into masks→ Crop & resize to 512×512→ 80/20 train/validation split→ Train two U-Nets→ Validate & apply to new images

Problem

Ultrasound is a cheap, non-invasive way to see muscles and tendons moving. But the images are noisy, results vary between researchers, and rules-based tools break when new data looks different. Even UltraTrack, the field's standard tool, needs manual correction and can't follow the long muscle fibers in the thigh.

Approach

I traced the top and bottom borders of the rectus femoris (a thigh muscle) in 490 images, then built two separate networks from Cronin's U-Net design. I imported the model into MATLAB, replaced its output with a pixel classification layer, and trained each network for 60 epochs on an RTX 2060 GPU using the Adam optimizer. Each took about an hour.

Broader lab work

Wrote an award-winning undergraduate thesis on ML-enhanced post-processing for real-world biomechanical signals
Combined IR motion capture data with ML to improve safety assessment accuracy by 42% in an exoskeleton study

Raw thigh ultrasound images next to model predictions highlighting the top and bottom muscle borders in cyan — Raw ultrasound image (left) next to the model's predicted top and bottom borders (right).

Border	Global acc.	Mean IoU	Weighted IoU	Mean BF
Top	0.9952	0.9157	0.9907	0.9926
Bottom	0.9907	0.8465	0.9825	0.9447

Research proposal · GTMI 455, Sept 2021 Open PDF in new tab ↗

Full report · PURA Final Report, May 2022 Open PDF in new tab ↗

Grand Challenges: NeuroPrep

Aug 2020 – May 2021

Team Member · Grand Challenges, Georgia Tech · 5-person team

NeuroPrep is a concept app that helps neurodivergent young adults build interview skills and social confidence for the job search.

Designed the app's features: finding mentors, events and workshops, building strengths, eye contact practice, managing anxiety, and applying to open roles
Created a step-by-step process to help unemployed autistic adults strengthen their job applications
Interviewed and surveyed 10 mental health professionals to ground the team's research
Positioned NeuroPrep against tools like Pramp, Proloquo2Go, and Simugator as both personalized and affordable
Pitched the app at the Grand Challenges Demo Day

1 in 68Children diagnosed with autism

53%Of autistic young adults unemployed after high school

50,000Autistic teens turning 18 every year

52%Of people the team surveyed wanted help with interviews

The NeuroPrep website.

Pitch slide mapping the app's six main features along a winding road — Six main features, from mentors to open roles.

Pitch slide showing four phone mockups of the app's landing, registration, sign-in, and dashboard screens — Landing, sign-up, and dashboard screens.

Pitch slide showing the eye contact practice exercise across three phone screens — Eye contact practice exercise.

NeuroPrep website ↗ Demo Day pitch deck ↗

Pathology Dynamics Lab

May 2018 – May 2021

Data Scientist · Georgia Tech

Found 3 clinical trends in cancer patients using unsupervised ML, including K-means and DBSCAN clustering
Text-mined 56 PubMed articles and curated the extracted patient data into a database

Peer-reviewed publication · Co-author

Meta-Analysis of Gastrointestinal Adverse Events from Tyrosine Kinase Inhibitors for Chronic Myeloid Leukemia

Mohanavelu P, Mutnick M, Mehra N, White B, Kudrimoti S, Hernandez Kluesner K, Chen X, Nguyen T, Horlander E, Thenot H, Kota V, Mitchell CS. Cancers (Basel). 2021;13(7):1643.

Tyrosine kinase inhibitors (TKIs) are the first-line drugs for chronic myeloid leukemia, but their digestive side effects affect patients' quality of life. The study pooled 43 studies covering 10,769 patients to compare how often four TKIs cause nausea, vomiting, and diarrhea. Unsupervised clustering showed that treatment outcomes depend mostly on how severe the disease is, not on which TKI is used. Doctors can therefore pick a TKI based on its side effects and the patient's health and lifestyle.

PubMed ↗ Full text (PMC) ↗ DOI ↗

Share of patients with digestive side effects, by TKI

Bosutinib

52.9%

Imatinib

24.2%

Dasatinib

20.4%

Nilotinib

9.1%

Differences between drugs were statistically significant (p < 0.001). Across all four, diarrhea was most common (22.5%), then nausea (20.6%) and vomiting (12.9%).

Mentra Startup Project

Aug 2020 – Jan 2021

Data Scientist · Data Science @ GT

Helped Mentra, a startup that connects neurodivergent job seekers with careers suited to their strengths, improve its job matching
Built NLP matching algorithms in Python that paired 100 neurodivergent clients with employers from a test database
Designed 20 ranking features to score candidates on skills, experience, education, and job interests
Grouped job listings with K-means clustering on TF-IDF text features, and mapped O*NET skills to relevant occupations

Mental Health & Wellbeing

Aug – Dec 2020

Undergraduate Researcher · Vertically Integrated Projects, Georgia Tech

Collected posts from Georgia Tech's subreddit by scraping Reddit with the Python Reddit API Wrapper (PRAW)
Measured emotions tied to anxiety and depression with sentiment analysis, and visualized them with word clouds
Logged weekly individual and team progress in a shared research journal

Word cloud of frequent words in Georgia Tech subreddit posts, with mad, nice, and doesn't among the largest — Word cloud from Georgia Tech subreddit posts.

Second word cloud from Georgia Tech subreddit posts, with use, basis, and students among the largest words — A second sample, highlighting school and campus themes.

Projects & experience

Mechanical Engineering

CAD, FEA, and hardware — from a patent-pending medical device to a sensor mount for a robot dog.

Senior Design Capstone

MediFlex Bell: a nurse call bell that adapts to the patient

About 5.4 million people in the US live with paralysis or limited mobility, yet 43% of nurse call systems still rely on a standard push-button. Hospitals buy a separate specialty cord for each patient's needs. We designed one modular bell that adapts to each patient instead.

Role: Founding Mechanical Engineer
Team: 5 engineers · Team Mechatronics
Advisor: Dr. Tequila Harris, Georgia Tech
Outcome: US provisional patent filed, Apr 2024

4Swappable input mechanisms in one device

555×Safety factor on the incline-pad mechanism

3.9×10¹²Estimated fatigue-life cycles, incline pad

18 VLow-voltage system on a custom 2-layer PCB

One bell, four ways to call a nurse

Breath callGooseneck-mounted mouthpiece with a pneumatic sensor, for patients with no hand mobility

Incline padA force-sensing tilted pad that activates with about 0.5 N of pressure

Soft buttonA silicone button with a trigger mechanism, for patients with limited grip

Remote buttonsA standard button pad for patients with typical hand function

From CAD to working prototype

CAD exploded view: removable casings, the four input modules, the PCB plate, and the rotating turntable base.

Finished MediFlex Bell prototype: a white 3D-printed cube with the gray-and-red soft button on the front, the incline pad folded out on the side, and the breath-call gooseneck mouthpiece on top — Final prototype: soft button, fold-out incline pad, and breath-call mouthpiece.

Another side of the prototype showing the remote button pad with raised blue and red buttons, and the breath-call gooseneck coiled in its side storage bay — Remote button pad, with the gooseneck stored in its side bay.

Objective

Design and prototype a specialty call bell that reliably connects nurses with limited-mobility patients, is configurable and easy to use, and plugs into a facility's existing nurse call system.

Design

The enclosure is about 5 × 5 inches, with removable casings, a rotating turntable base that stays connected to power as the bell is reconfigured, a built-in power cord path, and speaker and microphone outlets for two-way talk. Inside, a snap-fit base plate holds the PCB, with storage space for the gooseneck and electronics.

A custom two-layer PCB links the force and pneumatic sensors, microphone, and speaker to a microcontroller over I²C.

Material selection

In Ansys Granta EduPack, I built material indices to keep mass and cost low while staying tough, water-resistant, and latex-free. I compared PC, PET, and ABS, and chose ABS for the frame, casings, and turntable because it's cheap to make and holds up to hospital disinfectants. I picked silicone elastomer for the soft buttons and FR-4 for the PCB, which is flame-retardant and about 2.5× cheaper than the alternatives.

My contributions

Designed and iterated the modular assembly in CAD, integrating the gooseneck, turntable, and sensor housings
Ran material selection and FEA with mesh-convergence studies
Wrote the FMEA risk assessment and mapped the design to medical device standards

Structural analysis (FEA)

Component	Load case	Result
Incline pad	0.5 N repeated	FoS 555 · 3.89×10¹² cycles
Breath-call mouthpiece	1 psi internal	FoS 247 · 4.26×10¹¹ cycles
Casing	10 N axial	FoS 464
Bell frame	250–400 N compression	Yields at 380–400 N
Gooseneck (steel)	Axial pull	Yields at 80 N

Top FMEA risks

Failure mode	S	O	D	RPN
Incline plane deformation	8	5	4	160
Bell frame deformation	9	5	3	135
Mouthpiece lifespan	7	4	4	112
Soft button lifespan	7	4	4	112

Standards addressed

IEC 60601-1 — safety of electrical medical devices (stress analysis on the frame and mechanisms)
UL 1069 — reliability of nurse call systems
AAMI HE75 — usability and product life cycle
IEC 80601-2-49 — patient monitoring devices
NFPA 99 — fire and electrical safety (PCB and power supply placement)

Final capstone presentation · April 2024 Open PDF in new tab ↗

Invention Studio

2021 – 2024

Prototyping Instructor · Georgia Tech makerspace

Staffed Georgia Tech's student-run makerspace
Taught students to use equipment like 3D printers, laser cutters, and CNC machines to build their own projects

OXOS Medical

Aug – Dec 2022

Mechanical Engineering Intern · Atlanta, GA

Cut electromagnetic sensor tracking error in the company's flagship x-ray device by 24% through sensor position testing in Unity
Revised engineering specifications in Onshape to secure FDA clearance, avoiding a ~$30,000 loss
Sped up team workflow by 15% to hit quality assurance deadlines

Ecolab

May – Aug 2021

Software Engineering Intern · Alpharetta, GA

Built a material tracking database with macros, reducing inventory errors and speeding time to market by 33%
Wrote 23 verification and validation reports in Python for FDA and EU medical device compliance
Ran 4 feasibility tests and 7 design iterations on a medical-grade monitor drape

Agile Locomotion & Manipulation Lab

Jan – May 2021

Mechanical Engineer · Vertically Integrated Projects, Georgia Tech

Designed a visual-inertial-lidar sensor box that lets the lab's Unitree A1 robot dog map its surroundings and track its own position in real time (SLAM)
3D printed and tested 15 parts across 3 assemblies to validate the box's structural stability and sensor placement
Mounted the box to the robot's rail with T-slotted framing, using a hollow mount to cut weight without losing strength
Worked with the electrical and design leads to fit the electronics into one integrated system

Final CAD design of the SLAM sensor box: a closed, vented enclosure with the lidar sensor mounted on its lid, sitting on the robot's base plate — Final design: lidar mounted on the lid of a vented enclosure.

The final sensor box with its lid lifted, showing the electronics packed inside — Lid lifted to show the electronics inside.

The final sensor box without its lid, showing the internal electronics bay and cooling vents along the sides — Electronics bay with cooling vents along the sides.

Exploded CAD view separating the lidar lid, the enclosure, and the base plate — Exploded view: lid, enclosure, and base plate.

Earlier CAD iteration of the SLAM sensor box: a lidar sensor sits on top of an open enclosure holding a small onboard computer, mounted on an aluminum rail above the robot's base plate — Earlier iteration: lidar, onboard computer, rail, and mount.

CAD model of the sensor box enclosure resting on a length of T-slotted aluminum rail — Enclosure on aluminum rails that slide onto the A1, Cassie, and Mini Cheetah robots.

CAD model of the redesigned mount with hollowed-out pockets and screw holes — Hollow mount redesign that cuts weight while keeping strength.

Cross-section of T-slotted aluminum framing with a screw and nut seated in the top slot — T-slot attachment: screws run through the box bottom into the rail.

SLAMbox SolidWorks files ↗

3D Printing Side Projects

2020 – 2021

Personal projects

Designed and 3D printed a birdhouse feeder, then set it up outdoors on a deck railing
Printed a multi-part box with a Georgia Tech logo lid

Gray 3D-printed birdhouse with a round entrance hole and a feeding tray, mounted on a deck railing — Birdhouse feeder, front.

The 3D-printed birdhouse feeder on a deck railing with trees behind it — Set up on the deck railing.

The birdhouse feeder with its roof open, filled with birdseed — Roof open, filled with seed.

Multi-part box with a Georgia Tech lid.

Projects & experience

Business & Finance

Product growth, monetization, and market research.

Product Management

AI-driven growth and monetization at Capital One

Product work across the Capital One web and mobile apps and Capital One Shopping — running experiments, building data pipelines, and deploying AI tools that drive revenue.

Roles: Sr. Associate PM, Growth · Associate PM, Monetization
When: Aug 2024 – Present
Where: New York, NY

+4.23%Conversion lift from A/B testing

$8.34MAnnual present value from a re-evaluated customer group

$1.8M/moNew revenue from an AI upsell tool

$92MRevenue supported across 212 client meetings

Growth — Web & Mobile App

Led A/B testing work with product and design teams, lifting conversion rate by 4.23%
Added AI-driven analysis to daily monitoring pipelines to surface product suggestions and KPI shifts in real time
Used root cause analysis to re-evaluate a previously restricted customer group, adding $8.34M in annual present value

Monetization — Capital One Shopping

Deployed an algorithmic AI tool that informs upsell strategy in real time, adding $1.8M in monthly revenue
Led a cross-functional effort with product and engineering to bring category-level ETL data into the sales team's CRM
Delivered rate and flat-fee recommendations for 212 client meetings, enabling $92M in revenue

Demotu

2025 – Present

Angel Investor & Board Member

Invested early in Demotu and serve on its board

Equity Research Boutique @ GT

Jan 2023 – May 2024

Analyst

Researched UBER, LYFT, CART, and DASH using the Bloomberg Terminal
Built DCF models forecasting 5-year stock growth for UBER and CART, with 92% and 88% accuracy
Presented quarterly earnings updates and predictions for Uber and Instacart

Science Olympiad @ GT

2021 – 2024

Vice President · previously Finance Director

Raised $10,000 for the state tournament through department sponsorships and student organization grants
Co-hosted the first-ever flighted Georgia State Science Olympiad Competition, splitting it into a division for emerging schools and one for experienced schools
Organized Georgia Tech's first-ever Yellow Jacket Invitational Tournament

LinkedIn post ↗

Education

M.S. Computer ScienceGeorgia Institute of Technology

B.S. Mechanical EngineeringGeorgia Institute of Technology · Robotics & Automation, Minor in CS (AI)

Technical: Python, SQL, Swift, MATLAB, TensorFlow, PyTorch, NLP, LLMs (Claude Code / Codex), Optimizely, JIRA, Figma
Analytical: ETL creation, dashboarding, data product development, financial modeling, client-facing work, root cause analysis
Mechanical: CAD (Onshape, SolidWorks), Ansys Granta, FEA (Siemens NX), 3D printing, laser cutting, water jet, lathe, CNC