David Sorensen

(240) 855-4536 · david-p-sorensen@proton.me · linkedin.com/in/david-p-sorensen · github.com/david-p-sorensen

I'm an Economics, Mathematics, and Computer Science triple major at the University of Maryland. I've just finished a research internship at the Federal Reserve Board, and I'm currently writing two papers over my senior year. My interests are in the intersection of data science and economic research. I plan to graduate in May 2027 with the goal of either pursuing a Ph.D. in Economics or a career in data science.

Education

University of Maryland, College Park

B.S. in Economics and Mathematics
B.S. in Computer Science (Machine Learning Track)

GPA: 3.67

Relevant Coursework

Economics: Computational Macro (PhD), Advanced Econometrics, Micro & Macro Theory, Game Theory, Behavioral Econ
Mathematics: Math Camp (PhD), Multivariable Real Analysis, PDEs, Mathematical Statistics, Probability Theory
Computer Science: Deep Learning, AI, Data Visualization, NLP, Numerical Analysis, Applied Linear Algebra, Algorithms

August 2023 - May 2027

The Avalon School

GPA: 4.7

Valedictorian

September 2019 - May 2023

Experience

Economic Research Intern

Federal Reserve Board · Washington, DC

Assisted the Senior Associate Director of International Finance, Matteo Iacoviello, on his Geopolitical Risk Index, contributing to the validation and methodology sections of the research paper.

Deployed Llama-3.1-8B with 4-bit NF4 quantization on a single Tesla T4 GPU on AWS SageMaker; engineered a batched inference pipeline (left-padding, attention masking, batch-size tuning) and benchmarked per-article cost and throughput against the paper's GPT-4o-mini pipeline, scoring 90,000+ articles across four historical crisis windows
Replicated the paper's human-validation exercise across open-source and proprietary LLMs; identified an information asymmetry between human annotators and the LLM, then designed and ran controlled reruns on 1,050 human-labeled articles to categorize human-AI disagreements
Conducted QLoRA fine-tuning learning-curve experiments over nested train-test splits and multiple random seeds, with chain-of-thought and few-shot ablations, establishing that zero-shot inference with the base model outperformed all fine-tuned configurations
Built a two-timescale steady-state Kalman filter estimating latent daily country-level risk from irregular news coverage, deriving precision-weighted gains from Fisher information and calibrating decay rates to hand-labeled conflict episodes

May 2026 - August 2026

IT Student Worker

University of Maryland · College Park, MD

Resolved 2,500+ Jira tickets across device onboarding, network diagnostics, printer support, and event assistance
Delivered real-time assistance through walk-ins and phone consultations, communicating with clients while troubleshooting under time pressure
Produced instructional videos and website updates to improve self-service resources for students and staff

June 2024 - May 2025

AI Software Developer Intern

Intellectmap · Rockville, MD

Transitioned the company chatbot from the OpenAI API to an open-source large language model, enhancing privacy and cutting API fees
Embedded company-specific data into the chatbot using retrieval-augmented generation
Created an automated mapping system using extensive C# and Excel functions to convert geographic data into coordinates on customizable map projections

August 2023 - June 2024

Math Tutor

The Avalon School · Wheaton, MD

Tutored over 50 high-schoolers, developing their academic confidence and enabling them to pass their courses

July 2021 - July 2023

Research

The Household Rate-Exposure Channel of Monetary Policy

Honors Thesis · Advisor: Borağan Aruoba · Current Draft

Measuring who bore the 2022-23 rise in credit card interest costs relative to household income and liquid buffers, using a shift-share exposure measure interacting pre-determined balances with prime rate changes in the PSID and SCF
Decomposing each household's balance-sheet exposure to the tightening (card interest, deposit interest, mortgage lock-in, inflation erosion of nominal debt, basket inflation) into sufficient statistics (Auclert, 2019) from PSID and SCF balance sheets, with Rubin's rules and replicate weights for SCF implicates
Constructing household-specific inflation rates from 31 PSID expenditure components mapped to category CPI series and a regularized cross-fitted permanent-income index to split incidence by liquidity, permanent income, and mortgage tenure
Running pre-specified power diagnostics, then reporting double/debiased machine learning and causal forest (grf) estimates in R as bounds with honest CIs after a zero-dose placebo exposed an interview-timing effect
Building a reproducible R pipeline (renv, Quarto) that links eight PSID family files (2009-2023) into a household panel, with a joint two-stage bootstrap for generated-regressor inference and Adão-Kolesár-Morales shift-share standard errors

Optimized Keyword Retrieval for Two-Stage LLM Classification Pipelines

Current Draft

Lead author of a methods paper with two Federal Reserve Board coauthors; building a Python tool that compiles an LLM yes/no classification prompt into an optimized Boolean keyword query executable on SEC EDGAR full-text search
Formalizing retrieval as cardinality- and budget-constrained maximum coverage (NP-hard, monotone submodular); deriving finite-sample uniform deviation bounds and characterizing three out-of-sample failure modes (optimism, budget violation, saturation) with a corrective for each
Building the solver: variance-regularized log-odds screening, greedy/GRASP construction, 1-for-1 swap local search, and iterated local search under lexicographic acceptance on a depth-weighted submodular objective
Designing the evaluation on a firm-year panel of 10-K/10-Q filings labeled by an LLM under three prompts (tariff exposure, refinancing pressure, labor constraints), with firm-level (CIK) cross-fitting and filer-clustered standard errors to prevent boilerplate leakage, and temporal splits to measure distribution shift
Engineering a reproducible, rate-limited EDGAR ingestion pipeline (quarterly-index sampling frame, SHA-256 replication manifest, no redistributed text) and validating emitted queries against the live EFTS API for realized coverage

Leadership

College Fed Challenge

Director of the University of Maryland's College Fed Challenge team, competing in 2025 and 2026 in a Federal Reserve-sponsored undergraduate competition simulating FOMC deliberations
Lead a five-person presenting team and roughly 40 student researchers across four research groups
Played the role of FOMC Chair, presenting the team's thesis, forecast, and policy recommendation
Estimated the effects of international shocks on U.S. inflation and activity, including the spread of a Middle East energy-price shock across inflation categories and tariff pass-through to consumer prices, alongside analysis of terms of trade, option-adjusted spreads, labor productivity, and capacity utilization

Projects

Location Quotient Mapping with LinkedIn Job Data

Developed an automated data pipeline using Python and the OpenAI API to process and classify 124,000+ jobs posted on LinkedIn, achieving granular industry and ownership categorization across all U.S. states
Engineered and maintained a comprehensive Excel dataset containing 9,000+ data points tracking state-level employment metrics and industry concentration parameters
Built an interactive web application in R using Shiny and Plotly libraries to create dynamic heat maps with custom color scaling and hover-based state analytics

Building and Hosting this Website through AWS

Built a static website with HTML, CSS, and JavaScript, hosting it with an Amazon S3 bucket
Deployed the website as a CloudFront distribution, using an HTTPS protocol, and created an alternate domain name using Route 53
Created a DynamoDB table to track the number of website views which gets updated with a Lambda function written in Python and is displayed on the website using JavaScript
Implemented Infrastructure as Code with Terraform for programmatic access to the AWS console
Maintain CI/CD by pushing changes to the website with GitHub

Azure Sentinel Honeypot SIEM

Used custom PowerShell script to extract metadata from Windows Event Viewer to be forwarded to third party API that derives geolocation data
Configured Log Analytics Workspace in Azure to ingest custom logs containing geographic info
Configured Custom Fields in Log Analytics Workspace with the intent of mapping this data in Sentinel
Configured Azure Sentinel (Microsoft's cloud SIEM) workbook to display global attack data (RDP brute force) on world map according to physical location and magnitude of attacks

Active Directory Home Lab

Simulated a Windows networking environment used by organizations by housing Active Directory on a virtual machine as a domain controller
Created two network adaptors on domain controller for the internet and internal clients
Assigned IP addressing, configured NAT, routing, and DHCP for the internal server
Created 1,000 users on the internal server with a PowerShell script
Connected another virtual machine to the internal server and logged in as one of the users

Skills

Programming Languages

Python, R, MATLAB (Dynare, OccBin), Stata, EViews, SQL, LaTeX, C, Java, Rust, C#, HTML/CSS/JavaScript, x86 Assembly, OCaml, Prolog, Ruby, PowerShell

Developer Tools

AWS (SageMaker, S3, CloudFront, Lambda, DynamoDB, Route 53), Azure, Terraform, Git, renv, Quarto, Jupyter, VS Code, Eclipse, Jira, Excel

Libraries

Pandas, NumPy, Matplotlib, Seaborn, Scikit-learn, PyTorch, Hugging Face Transformers, PEFT/QLoRA, bitsandbytes, grf, ggplot2, Shiny, Plotly, OpenAI API