Selected work
The shipped artifacts.
Three projects I lead with — one in production at a Fortune-100 telecom, one personal end-to-end build, one academic RAG capstone — plus production work and systems coursework. Each card opens a deep-dive.
T-Mobile for Business · Enidus
01 / 08
AI Chatbot & Agentic Copilot
Production agentic copilot for T-Mobile for Business with hybrid retrieval design (Qdrant vector + BM25 + RRF), durable SQL-backed flow state, and embedding-cache cost engineering. Stage-and-confirm UX so the LLM never autonomously writes.
Production agentic copilot at T-Mobile for Business — pilot with 15 reseller tenants and 100+ portal users across 25+ enterprise customers.
Verified Error-Bounded LLM Caching Gateway (Go) · Personal
02 / 08
semblance
An OpenAI-compatible LLM gateway in Go that adds verified semantic caching: it holds the rate of wrong cached answers under an error budget you choose, where a static similarity threshold cannot. A clean-room build of the vCache algorithm (ICLR 2026), deployed on Kubernetes.
An OpenAI-compatible LLM gateway in Go implementing verified error-bounded semantic caching, a clean-room build of vCache (Schroeder et al., ICLR 2026); deployed on Kubernetes (k3s) via a Helm chart.
Reliability & Safety Harness for LLM Cloud Agents · Personal
03 / 08
CloudGuard
Test-driven reliability and safety harness for LLM agents that operate cloud infrastructure. Measures tool-selection accuracy, blast-radius guardrails, and prompt-injection resistance against a real AWS mock (Moto), with every headline number written to a reproducible JSON artifact.
Reliability and safety harness for LLM agents that run cloud infrastructure, measured against a real AWS mock (Moto), test-first with 57 tests.
Local vs Hosted LLM Non-Inferiority Study · Research
04 / 08
GoodEnough
A preregistered non-inferiority study asking whether a quantized 1.7B model on a consumer CPU can replace a hosted 70B baseline. It cannot: the hypothesis was rejected on every slice. The result worth carrying is the bound it puts on how much any router could recover.
Preregistered non-inferiority study: can a quantized 1.7B local model on a consumer CPU stay within a 10-point quality margin of a hosted 70B baseline? The answer is no, on every slice tested.
Full-Stack RAG · UW Madison
05 / 08
Denari RAG Capstone
Senior capstone. Hybrid retrieval (BM25 + TF-IDF) over 22K+ documents and 300K+ embeddings — 73% QA accuracy, 40% latency reduction.
Full-stack RAG system over 22K+ documents and 300K+ embeddings, built as my UW Madison senior capstone.
Self-Serve Analytics Platform · Enidus
06 / 08
Custom Reports & Dashboards
Full-stack reporting + custom dashboards product I built end-to-end at Enidus. React + Node + SQL Server, multi-tenant, scheduled email delivery, visual query builder, hardened against SQL injection at every layer.
Self-built full-stack platform letting enterprise customers compose reports, custom dashboards, and charts over their own data — no engineering tickets, no SQL knowledge required.
Agentic Resume Tailoring Pipeline · Personal
07 / 08
ClaudeJob
End-to-end solo build. Node + Anthropic SDK pipeline that powers my own AI Engineer applications — and produced this site.
Personal Node + Anthropic SDK pipeline that ingests live JDs and tailors a structured-output JSON resume per role.
Distributed Systems · UW Madison
08 / 08
Real-Time Weather Pipeline
Cassandra + Spark pipeline ingesting time-series data from 1300+ weather stations. Built for write throughput and fault tolerance.
Distributed-systems coursework at UW Madison — Cassandra + Spark pipeline for 1300+ weather stations.