Portfolio • GitHub • LinkedIn • Email

A hybrid music recommendation system that finds songs with a similar feel by combining cultural retrieval, metadata matching, audio embeddings, vector search, and LLM-generated explanations.

• Architecture: Pulls candidates from Last.fm/ListenBrainz, matches tracks against MusicBrainz, and reranks by audio similarity using LAION-CLAP embeddings with PostgreSQL + pgvector.
• Performance: Designed a lazy, self-growing vector corpus that embeds only candidates surfaced by real queries, reducing repeat-run latency by roughly 60×.
• Evaluation: Built a 19-seed evaluation harness across 8 genres to compare CLAP audio reranking against cultural-retrieval baselines and HNSW nearest-neighbor search.
• Product: Shipped a Next.js console for visualizing pipeline telemetry and explaining why recommendations were returned.

Stack: FastAPI, Python, PostgreSQL, pgvector/HNSW, Next.js, TypeScript, Docker, LAION-CLAP, Anthropic Claude

A full-stack Yu-Gi-Oh collection and pricing system that treats a trading card collection like an investment portfolio: cost basis, market value, reconciliation confidence, pricing coverage, and unrealized gain/loss.

• Data engine: Built a confidence-scored CSV import and reconciliation pipeline that matches collection exports against a 43,000+ printing catalog while preventing cross-edition pricing errors.
• Portfolio logic: Tracks per-lot cost basis, market value, and unrealized gain/loss instead of flattening ownership into a single average price.
• Infrastructure: Self-hosted on a Hetzner Linux VPS behind Caddy TLS with containerized services, pull-based deployment, nightly systemd ETL jobs, and off-box PostgreSQL backups to Cloudflare R2.
• Quality: Shipped a Next.js product UI with interactive valuation views, OpenAPI-generated TypeScript clients, 873 automated pytest/Vitest/Playwright tests, and a live ops dashboard for pipeline health.

Stack: Django 5.2, Django REST Framework, PostgreSQL, Next.js, React, TypeScript, Playwright, Docker, Caddy, Cloudflare R2

I currently work as a Software Engineer at CelLink, where I build and maintain internal production software used across multi-site manufacturing operations.

Selected work includes:

• LLM agent: Built a production Microsoft Teams assistant using FastAPI, the Anthropic SDK, and the M365 Agents SDK, enabling 500+ users to query live SQL Server data in natural language and cutting ad-hoc data requests by 80% at roughly $115/month.
• RAG system: Built an internal developer Q&A system over 170+ Confluence pages using semantic vector search, locally hosted LLMs, and the Atlassian REST API, reducing recurring software-team support requests by 20%.
• Internal platforms: Architected and maintained 15+ production ASP.NET MVC applications on SQL Server serving 500+ employees across California and Texas sites with 99.9% uptime.
• Performance: Reworked critical SQL Server procedures with cursor elimination, set-based CTE/CROSS APPLY rewrites, and composite indexing, reducing p95 execution time from 11s to 500ms.
• Data infrastructure: Designed 20+ event-driven ETL pipelines processing 50,000+ manufacturing records daily, then migrated them to a K3s Kubernetes cluster with multi-node failover.

I like building pragmatic systems that survive contact with real users: clean data models, observable pipelines, boring deployment paths, fast feedback loops, and UIs that make messy operational data easier to trust.

• Software Engineer at CelLink
• B.S. Computer Science, University of Minnesota — Twin Cities
• Minors in Information Technology and Digital Media Studies
• Experience across AI tooling, internal platforms, manufacturing software, ETL, full-stack applications, and self-hosted deployment