Matthew Kuan

Python · Google Gemini API · Decimal · pytest

HackerRank Orchestrate, Sep 2026 · #56 / 3062

Placed #56 of 3,062 with an affordability agent where the model reads evidence but never produces a number.

• Built a financial agent that decides whether a user can safely afford a purchase and recommends paying in full, partially, in installments, waiting, or not proceeding, across 250 requests in five currencies.
• Reconciled each user's transactions into a ledger (cancellations, duplicates, linked lifecycles, dated FX, recurring series detection) and projected a 90-day daily balance to test every plan against their minimum balance.
• Solved the largest safe payment with a cent-level binary search and the earliest safe full-payment date with a forecast scan, then ranked candidate plans on six ordered rules to set status and payment method.
• Limited Gemini to classifying messages and reading amounts off images through closed-schema JSON, with prompt-injection flagging, a numeric guard on explanations, a response cache, and a scoring harness over golden fixtures.

Angular · Spring Boot · Java · SQLite · Ollama · whisper.cpp · Kokoro TTS

Personal Project

Turns lecture PDFs, recordings, and YouTube links into notes, flashcards, and quizzes, fully on-device.

• Built a local-first study tool: drop in lecture PDFs, audio, video, or a YouTube link and get study notes, key terms, flashcards, and a self-grading quiz, with no data leaving the machine.
• Wrote a Spring Boot backend that extracts and cleans PDF text page by page, transcribes audio locally with whisper.cpp when captions are missing, and runs map/reduce generation over long documents with Ollama.
• Added a Listen mode that reads notes aloud with browser voices or a Kokoro sidecar, rewrites math into spoken words first, and highlights the passage being read so clicking any section jumps the audio there.
• Packaged the Angular front end into a single runnable jar with one-click installers for the optional speech and media tools.

Java · Spring Boot · Angular · PostgreSQL · Python · Gemini API

Emerging Leaders Series Capstone, 2026

Full-stack return model over live market data, covering 48 mutual funds and 20 bond funds.

• Built RESTful APIs that project investment returns from historical performance, U.S. Treasury rates, and fund-level risk factors, then surface expected value and risk profile through an Angular front end.
• Integrated Yahoo Finance for 5-year historical returns, the FRED Treasury API for risk-free rates and term/credit factors, and Newton Analytics for beta, caching live values in Postgres to cut redundant API calls.
• Seeded bond fund regression coefficients with a Python job, and added a Gemini tool-calling chatbot that fetches live expected returns and future value projections under explicit recommendation guardrails.
• Covered the math with JUnit tests asserting that future-value computations return the correct number of periods and stay positive and finite across every fund category.

Next.js · TypeScript · Supabase · Stripe · Expo (React Native) · CSS

Personal Project

A live, revenue-ready product: Stripe billing, Supabase backend, and a React Native app.

• Built a full-stack sleep-wellness app that quantifies sleep debt on a rolling 14-day ledger, estimates a personalized melatonin wind-down window from logged habits, and correlates daily energy ratings against accumulated debt.
• Implemented a freemium model with Stripe subscription billing (checkout, webhooks, and billing portal) over a Supabase Postgres backend with authentication and SQL migrations.
• Shipped a companion Expo / React Native mobile app with automated over-the-air deployments via an EAS Update GitHub Actions pipeline triggered on each push.

Gemini API SDK · PyYAML · Pillow · tenacity · tqdm

HackerRank Orchestrate 2026 · #85 / 1773

Placed #85 of 1,773 by cutting per-claim cost to ~$0.001 with a two-stage vision pipeline.

• Built a two-stage multimodal pipeline (Google Gemini vision) to adjudicate insurance-style damage claims across cars, laptops, and packages, classifying image evidence as supporting, contradicting, or insufficient.
• Designed a cheap Stage-1 validity gate (Flash-Lite) that screens for blur, wrong-object, and quality issues before triggering expensive Stage-2 analysis, cutting cost to ~$0.001/claim while improving flag reliability.
• Engineered object-specific prompts with injected evidence rubrics and user-history risk context, plus a closed-vocabulary normalizer and evaluation harness (macro-F1, confusion matrix) scored against labeled data.

React · Vite · Node.js · Express · PostgreSQL · OpenAI

Healthcare Innovation Challenge 2026

AI recovery-support app with encrypted health data, pitched to a panel of healthcare judges.

• Built a full-stack web app using AI to personalize daily recovery micro-tasks for users with substance use disorder, converting mood check-ins and journal entries into actionable interventions via the OpenAI API.
• Designed a relapse-risk classification system that analyzes mood, cravings, sleep quality, and journal sentiment in real time to surface targeted recovery actions.
• Implemented JWT-based authentication and a RESTful Express/PostgreSQL backend with row-level data security to protect sensitive health information.
• Presented to a panel of healthcare judges, competing on feasibility, scientific rigor, and public health impact.

Next.js · React · Tailwind CSS · Google Gemini API

L’Oréal Brandstorm 2026

AI fragrance recommender with a circular-packaging model, presented at L’Oréal Brandstorm.

• Developed an AI-powered personalized fragrance recommendation web app integrating the Google Gemini API to analyze user outfit images, local weather, and personal preferences, delivering scent recommendations and application guidance.
• Designed a multi-step discovery journey (location, climate, personality, and sillage inputs) using Next.js and Tailwind CSS to deliver a premium, bespoke digital consultation experience.
• Conceptualized a circular business model featuring refillable aluminum-shell fragrance vessels and a seasonal scent cadence using PCR glass (reducing CO₂ up to 30%), aligning with L’Oréal’s 2030 sustainability goals.
• Presented product strategy and tech architecture to panel judges, positioning Savoir-Faire as a “beauty tech” brand combining AI-driven personalization with sustainability-first hardware design.

Next.js · MongoDB · Docker · NextAuth · Demucs · basic-pitch

HopperHacks WiCS 2026

Real-time pitch detection and 3D tab rendering in the browser, shipped at a hackathon.

• Built a gamified, full-stack guitar practice platform with real-time pitch detection, MIDI file import, and 3D tab visualization using Three.js and React Three Fiber, enabling users to practice along with animated guitar notation.
• Implemented real-time audio processing using the Web Audio API and basic-pitch for pitch accuracy feedback; integrated Demucs for audio stem separation via a Dockerized Python backend, enabling dependency-free client-side usage.
• Developed user authentication (NextAuth), profile management, campaign modes, and a competitive leaderboard using MongoDB, fostering a social and structured practice ecosystem.
• Containerized the audio import service with Docker, pre-warming the Demucs model at startup and caching checkpoints in a persistent volume, to eliminate manual dependency installation for contributors.

Fusion 360 · Arduino · C++ · Python · MediaPipe

Emergent Biodesign 2026

Hands-free robotic control from webcam hand tracking, with per-joint safety validation.

• Designed and 3D printed custom human-finger-inspired gripper modules in Fusion 360 and Meshmixer, resolving repeated print failures caused by filament instability and structural fracturing through iterative redesign.
• Programmed Arduino firmware in C++ to initialize the Braccio arm to a calibrated safe position, parse six-joint angle commands from serial input, validate each value against a safety table, and execute incremental 3° motion steps with inter-step delays to prevent mechanical strain.
• Built a Python vision-control pipeline using MediaPipe to capture real-time webcam feed, identify bilateral hand landmarks, and map right-hand pose to arm joint angles and left-hand pose to gripper open/close commands.
• Benchmarked design against four existing control paradigms; differentiated solution by eliminating pre-programmed motion sequences and preserving continuous surgeon judgment in the control loop.
• Delivered a 10-minute live in-class demonstration of real-time surgical tool grasping, arm calibration, and safe reset; presented circuit schematics, firmware architecture, and vision pipeline to peer and faculty reviewers.

Fusion 360 · AutoCAD

Emergent Biodesign, Spring 2026

Accessible one-push dispenser for users with limited hand mobility, toleranced to 0.15 mm.

• Designed a spur gear-based dispensing mechanism for wheelchair users with limited hand mobility, applying kinematic calculations to determine pitch diameter (4.0 in), diametral pitch (6), and tooth count (24).
• Produced annotated orthographic projections, assembly drawings, and motion simulations in Fusion 360; applied interference analysis to identify collisions and used offset face tolerancing to establish a 0.15 mm clearance fit.
• Ran parametric modeling with driven dimensions to dynamically link treat size constraints to gear geometry, enabling automatic recalculation of housing clearances when design parameters changed.

Research Writing · Energy Policy · Sustainability Analysis

Intermediate Writing (WRT 102), Spring 2025

Research assessing where AI genuinely reduces energy demand versus where it adds to it.

• Authored a research-based argumentative paper evaluating whether artificial intelligence can reduce net environmental harm, weighing the energy cost of AI runtime against its capacity to cut global consumption.
• Analyzed AI’s role in siting and scaling renewable energy, including offshore wind capacity assessment, and in operating energy plants more efficiently, drawing on peer-reviewed engineering and environmental sources.
• Evaluated real siting constraints for renewable installations, including average wind speeds, geotechnical foundation conditions, coastal biodiversity impact, and project economics, then peer-reviewed essays across other disciplines to refine argumentation and structure.

Python · NumPy · Matplotlib · Pandas

Personal Project

Dimensionality reduction and clustering pipeline for pattern recognition on complex datasets.

• Interpreted datasets using advanced computational methodologies, implementing DBSCAN and K-means clustering algorithms with heat-mapping visualizations to perform pattern recognition and validate data accuracy.
• Developed a Principal Component Analysis (PCA) program to streamline data interpretation and accurately isolate significant data regions.

C++ · Arduino · Fusion 360

Intro to Biomedical Engineering, Fall 2024

Led a 7-person team to a working wearable prototype; won Best Presentation.

• Authored technical documentation and delivered project presentations, earning the “Best Presentation Award” for exceptional communication, critical thinking, and clearly explaining the design process to non-technical audiences.
• Directed a 7-person team through an iterative project lifecycle, researching hypoallergenic, cost-efficient design materials, project requirements, constraints, and product design while managing timelines.
• Designed, assembled, and systematically troubleshot electronic hardware components to prototype a noninvasive wearable sensor, actively evaluating factors that impacted product performance, safety, and manufacturability.

Synthetic Biology · Recombinant DNA Design · SnapGene

BME 304 / BME 300 · Stony Brook University

Engineered a probiotic delivery route for tumor neoantigens, grounded in Nature literature.

• Investigated engineering probiotic E. coli Nissle 1917 within the gut microbiome to deliver tumor-specific neoantigens and enhance cancer immunotherapy outcomes, working in a team of six.
• Modeled recombinant plasmid construction incorporating tumor-targeting vectors, a pBAD inducible promoter, and listeriolysin O sequences; authored an accompanying critical analysis grounded in peer-reviewed Nature literature.

INSTRON UTM · IBM SPSS · Statistical Analysis

BME 212 · Stony Brook University

Measured a statistically significant 15% drop in bone force-to-failure (p = 0.009).

• Designed and executed an ex vivo three-point bending study on 20 matched-pair specimens to quantify how acidic pH affects bone structural integrity, conducting a power analysis to justify sample size.
• Applied paired t-test analysis in IBM SPSS to establish a statistically significant 15% reduction in force-to-failure (p = 0.009); documented methodology and findings in a formal paper and group presentation.

Fusion 360 · AutoCAD · EEG Concept Design

ART × BME Interdisciplinary Design 2026

Wearable EEG concept for pediatric narcolepsy, co-designed with a fine arts student.

• Collaborated with a fine arts student to prototype a wearable EEG-based pediatric narcolepsy monitoring headband in Fusion 360, iterating on a non-circular 20 cm profile to resolve sizing and geometry failures across print cycles.
• Proposed integration of EEG electrodes, signal amplifiers, Bluetooth modules, and onboard storage for real-time brainwave monitoring and wireless transmission to a caregiver companion app.