Things I’ve built

A volunteer-nonprofit matching system.

• Cleaned and integrated 4,600 nonprofit records from IRS, NCCS, DataSF, and the California Registry of Charities into a curated set of 1,500 organizations.
• Used Claude to write mission statements for organizations with incomplete descriptions, prompting the model to search their websites and summarize them.
• Deployed the team’s BERT-based recommendation model to an AWS SageMaker serverless endpoint and integrated it with backend services.

• BERT
• AWS SageMaker
• Data integration

A hybrid deep learning model that names a poem’s form.

• Built a BERT-BiLSTM architecture in TensorFlow/Keras that classifies poems into nine forms, reaching a weighted-average F1 of 0.86.
• Assembled a 2,270-poem dataset from multiple sources, with cleaning, parsing, and metadata standardization.
• Used class weights to address the significant class imbalance while preserving all poems and their potentially unique linguistic patterns.

• TensorFlow
• BERT
• BiLSTM
• NLP

An analysis of California Highway Patrol crash records for Napa County, 2016–2026.

• Loaded statewide California Crash Reporting System files into PostgreSQL, then filtered, cleaned, and joined crash, party, and injury tables into an analysis view.
• Used window functions, CTEs, and CASE bucketing to analyze when crashes happen, how yearly totals changed, and how often someone under the influence was involved.
• Built a dashboard that reads the notebook’s exported results, so the charts and figures update whenever the analysis is rerun.

• PostgreSQL
• SQL
• Jupyter
• Plotly.js

An LLM pipeline that turns an RFP for audit services into a compliant proposal letter.

• Parses RFPs, drafts proposal letters with Claude, and reviews them for PCAOB independence issues in a LangGraph pipeline.
• Keeps drafting and reviewing as separate, testable steps; the reviewer flags problematic language, then retrieves PCAOB standards from a Chroma vector store to cite the relevant rule.
• Runs a bounded revision loop that produces a final letter, a structured change log, and a human-review status.

• Claude API
• LangGraph
• Chroma
• RAG

A RAG-based editorial chatbot that answers style questions from the Wikipedia Manual of Style.

• Built with AWS Bedrock (Claude), LangChain, and a Chroma vector database; deployed serverless on AWS.
• Designed and ran adversarial tests for prompt injection, role manipulation, and capability expansion, and documented the defenses and trade-offs.

• AWS Bedrock
• LangChain
• Chroma
• Security testing

A full-stack app that catalogs items from photos.

• Runs each photo through a Claude-powered LangGraph pipeline: six nodes across five stages (category routing, parallel identification and OCR, description, structured extraction, validation), with human-in-the-loop confirmation via graph interrupts and SSE streaming.
• Defines explicit node contracts against hallucination, including a three-state OCR output (text present / unreadable / none) that stops vision models from inventing label text.
• Gates the repo with automated PR review through Claude Code Action and real-Postgres integration tests; per-node evaluation against fixture images is underway.

• Expo / React Native
• FastAPI
• Supabase
• LangGraph

Writing and editing samples from ML research reports to technical communication coursework at UC Berkeley.

• A 33-page internal proposal for adding push notifications and dunning management to a fictional SaaS payments platform, with my own layout and Tableau visualizations.
• A substantive edit of a college website’s technical description of 3D printing, shown side by side with the original.
• A process description of how a hurricane forms at sea, written for general readers.
• Two statistical reports, on voting difficulty (ANES 2022 Pilot Study) and on leisure time and age (2022 American Time Use Survey). I co-wrote both and did the final edit.
• The research report behind my poetic form classification model.

• Technical writing
• Substantive editing
• Proposals
• Statistical reports