Caleb Bucci

A platform that helps homeowners find buildable lots, understand their constraints, and assemble the team needed to actually build.

• Architected the full-stack application in Next.js and FastAPI, backed by PostgreSQL with PostGIS for geospatial lot queries, deployed via Railway and Vercel.
• Built an agent pipeline with role-scoped tool access to automate parts of the software development workflow.
• Designed agents that convert acceptance criteria into frozen tests, and gate any compliance-affecting (sensitive) change behind human approval.

• FastAPI
• Next.js
• PostgreSQL
• Railway
• Vercel
• Agentic Workflow
• TypeScript

More than 1 in 3 American adults are classified as prediabetic. This project can screen for prediabetes using five features a patient can measure themselves. It can also predict continuous HbA1c levels if a patient has existing lab data.

• Trained an XGBoost regression on NHANES 2021-2023 dataset to predict continuous HbA1c, achieving a tolerance of 0.39%.
• Reframed the task as a binary screener for prediabetes (HbA1c ≥ 5.7), reaching 0.76 holdout ROC AUC.
• Applied SHAP attribution to surface the most influential biomarkers per patient; the top features aligned with established clinical literature.
• Served the model end-to-end to a SwiftUI iOS app through a FastAPI REST endpoint.

• XGBoost
• SHAP
• Python
• SwiftUI
• iOS
• Pandas
• Scikit-learn

A study to determine where extractive question-answering models fail, and whether targeted adversarial data can repair those failures and improve the model.

• Fine-tuned ELECTRA-small on SQuAD for extractive QA, then ran behavioral testing with CheckList across categories including negation, coreference, and temporal reasoning.
• Compared three fine-tuning strategies using LoRA adapters.
• Generated adversarial examples from the failed categories and AdvSQuAD, cutting the CheckList failure rate from 62% to 43%.

• Python
• Pytorch
• LoRA
• CheckList
• HuggingFace
• Scikit-learn

A mobile app for Android devices developed in Kotlin which allowed users to share their music listening habits with their friends as well as share music directly.

• Implemented Spotify OAuth 2.0 with PKCE, Firebase Firestore/Storage for shared multi-user state, and a media player where users could upload, download, and play each other's MP3s.
• Displays Spotify and local listening history on each user's profile.

• Kotlin
• Android
• Firebase
• Android SDK
• REST APIs
• Oauth

DataAnnotation | Remote | Contract/Part-Time

• Designed adversarial prompts in Python, Java, and JavaScript that expose code and logic errors in LLMs, improving training data quality.
• Wrote reference solutions and completed rubric-based evaluations of model output, feeding failures into RLHF pipelines.

• Python
• Java
• Javascript
• RLHF

Trimble Inc. | Dayton, Ohio | Internship

• Resolved a critical bug in Firebase user engagement tracking, reducing inflated metrics by 78.9%.
• Updated OEM attachment naming in Earthworks GO across both data storage and UI, in Swift and Android (Java).
• Developed a Python feature to automate license generation, streamlining App Store compliance.
• Optimized the translation process by removing over 3,000 redundant lines, cutting translation costs by 49%.

• Python
• Java
• Android
• Jira
• Jenkins
• C++
• Firebase
• iOS
• Swift

General Motors | Warren, Michigan | Internship

• Built a C test suite for an embedded display unit and rearview camera, validating video integrity and latency.
• Diagnosed camera glitches by injecting simulated temperature readings to test the firmware's overheating response.

• C
• Firmware Testing
• Embedded Systems
• GDB
• Jenkins
• Jira

University of Texas at Austin | GPA: 3.97

The Ohio State University | GPA: 3.85

Mathematics minor, AI specialization, Dean's List