Diya Vij

Bachelors in Applied Mathematics & Statistics; Minor in Data Science

Relevant coursework: Machine Learning, Bayesian Statistics, Time Series, Regression Analysis, Game Theory, Data Structures & Algorithms, Stochastic Processes, Data and AI Ethics

• Direct operations and marketing for Cal Hacks, the world's largest collegiate hackathon, and the UC Berkeley AI Hackathon; oversee a $1,250,000+ annual budget and 30+ organizers across logistics, sponsorships, and outreach.
• Scaled the applicant pipeline to 25,000+ registrants and 4,000+ hackers per cycle by owning top-of-funnel growth across channels; managed 50+ sponsor partners including Anthropic, Google, and Amazon.
• Designed and launched the org's first short-form video program (Instagram Reels + TikTok), personally producing the content that drove a step-change in reach and grew Instagram from 4,000 to 6,000+ followers.
• Piloted the organization's first Meta (Facebook/Instagram) paid-ads campaign, building the acquisition funnel from scratch.

• Shipped a two-path RAG pipeline (semantic vector search + keyword retrieval over PubMed, ClinicalTrials.gov, and UK Biobank) that turned a multi-day research task into a minutes-long query for a department of 50+ researchers.
• Engineered 600+ features from time-series data across 500+ patients and trained an XGBoost model reaching 85%+ accuracy detecting disease-severity progression; presented to medical directors and at a 150+ person symposium.

• Refactored SSIS ETL pipelines into Python and built Power BI dashboards for Facilities Financial Services, improving data-driven reporting across university IT systems.

• Built a multi-step agent (hand-rolled Claude tool-use loop, no framework) that reconstructs an argument's premise structure — including unstated "suppressed" premises — and separates genuine logical failures from the acceptable compression every real argument uses, grounding each flag in cited text and routing low-confidence calls to a human reviewer.
• Shipped a two-layer evaluation harness — academic fallacy benchmark (LOGIC/LOGICCLIMATE) for breadth + a hand-labeled set for depth — scoring detection recall, false-positive rate, evidence grounding, and confidence calibration (ECE/Brier); live token-streaming demo (SSE), deploying on Railway.

• Built a mobile-first AI journaling app: entries embedded via nomic-embed-text (pgvector), retrieved via hybrid semantic + recency scoring, queried through Claude Sonnet; a Supabase cron autonomously runs weekly synthesis via the Claude API.

• Built a menstrual-cycle mood tracker powered by a Bayesian model trained on the mcPHASES wearable dataset (PhysioNet); personalizes predictions as users log data and sends cycle insights via Twilio SMS.

• Co-built a CNN–Transformer hybrid classifying sleep apnea from 150 GB of multimodal EEG/ECG data (DREAMT), achieving 78% accuracy and 0.86 ROC-AUC across 4 model architectures.

• Berkeley Math Equity & Inclusion Committee rep.

• Part-time (10–20 hours/week) IT support for users at UC Facilities and UC Berkeley Haas.

• Managed club operations and alumni relations.