Hans Luther

Oval Brief turns U.S. executive orders into something citizens can actually read. I designed and built it solo — architecture, pipeline, back end, mobile app — from a blank repo to production in about five weeks part-time.

The back end is six Cloudflare Workers on a queue-based async pipeline over Neon PostgreSQL with pgvector. It processes 1,500+ executive orders from 1994 to today, producing plain-English summaries, calibrated 1-10 impact scores, and classifications across a 26-topic taxonomy and a 32-affected-group scheme. Claude handles analysis; OpenAI supplies 1536-dimensional embeddings.

Cross-references between orders are found in three phases: regex for explicit citations, embedding similarity for implicit ones, and then LLM semantic validation. Moving candidate selection from taxonomy matching to embeddings lifted precision from around 14% to over 50%.

The hardest problem was classification accuracy. Systematic prompt engineering against a 72-order calibration set cut the affected-groups error rate from 31% to under 1%.

I developed the mobile app in React Native/Expo for iOS and Android, tRPC for end-to-end type safety. It's built with agentic development using Claude Code: 962 commits and 221 PRs.

Anchor AI turned meeting recordings into actionable intelligence. I led engineering and designed the GenAI pipeline — built in spring 2023, before Claude had tool use, JSON mode, or structured output.

Audio flows through AssemblyAI for transcription, then Cloud Pub/Sub fans out to five parallel Cloud Functions running Claude via the LemUR API. Each meeting yields a one-sentence summary, a topical summary, paragraph-level minutes, and action items with speaker attribution — plus auto-generated titles and prep agendas built from the previous meeting's open items.

Getting structured data out of early Claude was the real work. Three parsing strategies: XML tags for simple values, line-prefix matching for topic summaries, and JSON coercion for action items — the last took 17+ regex iterations to survive inconsistent quoting and explanatory preamble. Model selection by task: Haiku for summarization speed and cost, Sonnet where reliable structured extraction mattered.

Results are written to Firestore and surface through real-time listeners, so the UI fills in as each feature completes. I also built "Max", a stateless Q&A assistant answering natural-language questions against full transcript context.

Fountain Life runs advanced diagnostics — Heart Scan, MRI, CCTA, CGM — across US centers. As a chief software development engineer, I designed the platform and led the 13 engineers building it.

The backbone is Google Pub/Sub, connecting 100+ microservices across App Engine, Cloud Run, and GKE, with a finite-state machine orchestrating 11 diagnostic journey types. Data sits in purpose-matched stores: BigQuery for analytics, MySQL for clinical records, Firestore for real-time state, Cloud Datastore for high-throughput documents.

The hardest problem was white-label delivery for 14+ partners without per-partner builds. I designed an iframe-based micro-frontend architecture: native iOS and Android apps render a React shell, while patient-facing screens are independently deployable App Engine services receiving auth tokens, patient data, and branding through URL parameters, with complex objects base64-encoded. Navigation is backend-driven — the state machine writes configs to Firebase RTDB and the shell re-renders reactively. Branding is entirely runtime via CSS custom properties, so one codebase serves every partner, and no native release is needed for content or branding changes.

ZeroBase was a real-time, multi-user financial planning platform for hospitality—built to replace Excel. As a founding engineer, I wrote the SOW, designed the business logic, and built it front to back.

The hard part was the calculation engine: an income statement hierarchy stored as individual Firestore documents, mixing rollup accounts, formula-driven accounts that reference other accounts, and labor accounts computed from position-level data.

Edits had to be completed in under 1 second, including the network round trip, which ruled out recalculating the changed branch—Firestore's per-document reads would pull dozens of untouched siblings at every level. I implemented incremental delta propagation: only the difference travels up to each ancestor, no sibling reads, no sums recomputed.

Full recalculation had no deterministic ordering—rollups depend on drivers, drivers on values, with cross-references throughout. I designed a bottom-up dependency resolver: calculate everything with no remaining dependencies, strip those from the others' lists, repeat. A topological sort across 3 account-type buckets, on Cloud Functions.

Angular, Akita, and RxJS up front. The MVP was shipped in under 3 months and won the first customer contract.

An application that consolidates data from regulatory reports such as FFIEC and NCUA from all banks nationwide for commercial real estate entities to identify lenders with potential sale assets.

I developed this app using Python to ETL the data into a Cloud Firestore database. A React front end accesses the data through a REST API exposed by Cloud Functions for Firebase.

I created this system as an alternative to the tracking architecture that required cellular service for deployment in remote event locations. I designed, assembled, and coded an ARM-based tracking device featuring a 900 MHz XBee radio and SiRF IV GPS receiver. The prototype was built with Tiva C LaunchPad.

The WPF client connects to the receiving XBee radio via a serial connection and plots GPS positions on a map enhanced with a predefined GPX course map.