Dylan Chen

Automating Neurodegenerative Disease Diagnostics

Momentous Health is building robotics-enabled systems to automate diagnostic testing for neurodegenerative diseases in elderly patients. The current workflow relies on manual test administration, paper-based result recording, and labor-intensive report generation—creating bottlenecks that delay patient care. I'm developing an agentic AI infrastructure that handles the entire clinical data lifecycle: from OCR-based extraction of test results, to automated report writing, to compliance verification before final delivery.

Report Writing Agent with Human-in-the-Loop

The core challenge is generating clinical reports that meet strict healthcare regulations while maintaining accuracy with patient health information. I built a multi-agent pipeline where OCR systems extract raw test data, reinforcement learning models refine extraction accuracy through human feedback, and RAG-powered language models pull the latest medical guidelines to generate compliant reports. Each agent specializes in one task—extraction, validation, or writing—then passes work downstream, creating a modular system that's easier to debug and improve than monolithic approaches.

PII-Aware Document Generation

Healthcare data requires careful handling of personally identifiable information. The system implements safeguards at every stage: OCR outputs are sanitized before storage, RAG retrieval filters out irrelevant patient data, and the final PDF generation layer applies role-based access controls. This architecture ensures compliance with HIPAA and other healthcare privacy standards while maintaining fast turnaround times for clinical staff.

From Manual Workflows to End-to-End Automation

The broader vision integrates this report writing system with robotics performing the actual diagnostic tests, creating a fully autonomous pipeline from patient assessment to results delivery. This work taught me how to design AI systems that operate in regulated, high-stakes environments where mistakes have real consequences—forcing careful attention to validation, interpretability, and failure modes that pure research projects often ignore.

Predicting Diabetes Complications Before They Surface

Type 2 diabetes affects over 6% of the global population, but current diagnostic tools only catch problems after clinical symptoms emerge—often when damage is already done. I'm working with the AI-READI dataset, a multimodal collection of 4,000 patient records balanced across race, ethnicity, gender, and disease severity, to build predictive models that identify risk trajectories hours to weeks before patients experience symptoms. The goal is to move from reactive treatment to proactive intervention.

Time-Series Forecasting on Longitudinal Health Data

The AI-READI dataset includes continuous glucose monitoring data, wearable sensor outputs, lab results, and environmental factors tracked over time. I'm implementing time-series forecasting models using JAX and GluonTS to analyze these longitudinal patterns and predict glucose spikes, hypoglycemic events, and disease progression markers. The challenge is handling data collected at varying sampling rates—CGM sensors log every 5 minutes while lab tests happen monthly—requiring models that can reason across different temporal resolutions.

Learning Patient-Specific Disease Dynamics

Rather than building one universal model, we're exploring whether individual patients exhibit unique disease signatures. The hypothesis: if you can fit a personalized model to each patient's historical data, the model's hyperparameters might reveal higher-order patterns that generalize across populations. For instance, two patients with similar glucose variability but different model learning rates might represent distinct disease subtypes with different intervention needs.

Engineering Multimodal Feature Pipelines

Raw sensor data isn't directly usable for ML models. I'm building feature extraction systems that process vitals, lab markers, and physiological signals into unified representations suitable for health prediction. This involves handling missing data (not all patients have all measurements), normalizing across different units and scales, and identifying which combinations of features best predict specific outcomes. The work sits at the intersection of signal processing, clinical domain knowledge, and machine learning architecture design.

Celebrating 25 Years with Cross-Campus Gameplay

CITRIS Quest commemorates the 25th anniversary of UC's Center for Information Technology Research in the Interest of Society by bringing students across four UC campuses together through a mobile game inspired by Space Invaders. Players physically explore their campuses to find and scan pixel art stickers, earning points for their campus in a semester-long competition. The game merges digital gameplay with physical exploration, requiring computer vision to work reliably under real-world conditions—varied lighting, camera angles, partial occlusions, and low-quality phone cameras.

Training CNNs for Real-World Robustness

I built a PyTorch CNN to classify pixel art characters from phone camera scans. The technical challenge wasn't accuracy on clean test sets—it was generalization to the messy conditions players would actually encounter: outdoor lighting, motion blur, tilted angles, shadows, and stickers that get worn or damaged over time. I implemented cross-entropy loss with learning rate scheduling and aggressive regularization (dropout, data augmentation) to prevent overfitting to pristine training images. The model needed to be both accurate and confident enough to reject non-game images without frustrating players with false negatives.

Full-Stack Product from Zero to Launch

Beyond the ML model, I led development of the entire product stack: Flutter mobile app for iOS and Android, Supabase PostgreSQL backend managing user authentication and scan storage, real-time leaderboards showing campus standings, and admin tools for monitoring player activity. I managed a cross-functional team through sprint planning, code reviews, and feature prioritization while balancing technical debt against shipping deadlines. This meant making architectural decisions early—like choosing Flutter over native development for faster iteration—and living with those constraints later.

Making AI Accessible to Non-Technical Users

The most valuable lesson was designing for actual user behavior, not ideal conditions. Early versions required perfect scans in good lighting. Real users took photos while walking, in shadows, or from weird angles because that's when they found stickers. I rebuilt the pipeline to handle this: multiple model passes at different confidence thresholds, preprocessing to normalize lighting and perspective, and UI feedback guiding users to better shots without being annoying. The goal wasn't just "build a CNN"—it was "build something students actually enjoy using."

Automating Failure Diagnosis at Supercomputing Scale

NASA's supercomputing infrastructure processes thousands of jobs daily across climate modeling, aerodynamics simulations, and mission-critical calculations. When jobs fail, engineers manually sift through logs to diagnose root causes—a bottleneck that delays research and wastes compute resources. ANDROMEDA automates this process using unsupervised machine learning to identify failure patterns, categorize issues, and surface interpretable insights without requiring labeled training data.

The Challenge: No Ground Truth Labels

Traditional supervised learning requires labeled examples of "this log pattern = memory overflow" or "this trace = network timeout." NASA's failure logs don't come pre-labeled, and manually tagging thousands of historical failures isn't feasible. The system needed to learn failure signatures directly from unlabeled data while producing explanations that domain experts could validate and act on.

Building Self-Supervised Pattern Recognition

I designed a contrastive learning architecture that treats similar failure traces as positive pairs and dissimilar ones as negatives, learning representations that cluster related failures together. The pipeline processes raw log data through semantic embedding models to capture contextual meaning, then applies HDBSCAN clustering to group failures by shared characteristics. An attention mechanism highlights which log segments drove each cluster assignment, giving operators transparency into why the system flagged specific patterns.

Production Integration and Interpretability

Rather than a black-box classifier, ANDROMEDA generates structured diagnostic reports mapping each failure to its closest historical matches and surfacing the distinguishing features. This lets engineers validate whether the system's pattern detection aligns with known failure modes. The architecture handles NASA's log volume and diversity while maintaining fast inference for real-time alerting.

Scaling RCA Without Manual Annotation

The system demonstrated that unsupervised learning could replace manual root cause analysis for recurring failure classes. By eliminating the labeling bottleneck, the approach generalizes to new failure types as they emerge rather than requiring retraining on annotated data. This work showed me how critical interpretability is when deploying ML in high-stakes infrastructure—domain experts won't trust opaque models, even if metrics look good.

Leadership Development at Berkeley Scale

Berkeley Student Leadership Academy is a year-long undergraduate program offering workshops on leadership, engagement, and diversity across campus. Students attend seven workshops minimum (four required, three elective) covering topics like team dynamics, communication strategies, and inclusive leadership. Managing workshop logistics—tracking attendance, coordinating facilitators, archiving feedback—generated massive amounts of data across hundreds of participants and dozens of events each year.

Replacing a Broken Spreadsheet Nightmare

The existing system was a single, bloated spreadsheet with hundreds of rows and columns filled with binary 0s and 1s indicating attendance, completion status, and requirements met. Finding information required scrolling through endless cells, updating records meant hunting for the right row, and setting up the tracker each semester took weeks of manual copying, formula adjustments, and data migration. The system worked barely, and only for people who already knew how to navigate it—a major barrier for new administrators.

Automated Tracker Setup in Under 10 Minutes

I built a Google Apps Script automation that generates the entire workshop tracking infrastructure from a template with a single button click. The script creates properly formatted spreadsheets with attendance tracking, automatically generates linked Google Forms for registration and feedback, populates formulas for calculating completion requirements, and sets up data validation rules to prevent entry errors. What previously took a couple weeks of tedious setup now happens in about 10 minutes, freeing administrators to focus on improving workshop content and participant experience.

Redesigning for Human Usability

Beyond automation, I restructured how data was organized. Instead of cryptic binary columns, the new system uses readable status indicators, color coding for quick visual scanning, and separate sheets for different data types (attendance vs. feedback vs. completion tracking). The archive system moved historical data into structured folders with consistent naming conventions, making it trivial to retrieve past program data for trend analysis or reporting. The goal wasn't just technical efficiency—it was making the system usable for future administrators without requiring institutional knowledge or extensive training.

Rethinking Phone Ownership Economics

Cyclick.ai is building a phone subscription model that challenges the traditional buy-and-own approach to personal electronics. Users rent phones instead of purchasing them outright, gaining flexibility to upgrade frequently, bundled repair coverage, and predictable monthly costs. When users return devices, Cyclick resells them to refurbishers in markets like Spain where there's strong demand for certified pre-owned phones. This circular business model reduces the company's capital requirements and passes savings to users through lower subscription prices.

Discovering Hidden Costs in Buy Now, Pay Later

Through user research and competitive analysis, I identified major pain points in existing BNPL (Buy Now, Pay Later) services: hidden fees buried in fine print, confusing interest calculations, and unexpected charges that destroyed trust. Users wanted ownership flexibility but were burned by predatory pricing structures. The insight: users weren't opposed to subscription models—they just didn't understand the economic benefits because competitors obscured the math. Our opportunity was radical transparency: show exactly what users save over time, compare total cost of ownership vs. subscription, and eliminate surprise fees entirely.

Competitive Pricing Strategy and Market Entry

I built pricing models scraping data from major BNPL competitors to understand their fee structures, interest rates, and total cost calculations. The analysis revealed a pricing band where Cyclick could undercut established players while maintaining healthy margins through efficient refurbishment partnerships. The key leverage point: while competitors treat returned phones as losses or sell them at commodity prices, Cyclick's Spanish refurbisher network paid premium rates for quality devices, creating additional revenue streams beyond subscription fees.

Expanding Accessibility in Emerging Markets

The subscription model's biggest advantage isn't just flexibility for affluent users—it's access for populations with limited purchasing power. Markets in Africa and other developing regions have strong demand for smartphones but lack capital for $800+ upfront purchases. Monthly subscriptions dramatically lower the entry barrier while Cyclick's refurbishment model keeps costs sustainable. I evaluated credit vetting integrations with Experian and Equifax to enable this expansion, balancing fraud risk against market accessibility.