Gunnar Hostetler

Why I built it: I was seeing a lot of discussions about the potential for local, on-device EHRs and patient privacy. I wanted to see if I could build a proof-of-concept myself, testing if the offline RAG engine I built for OpenIntelligence could be adapted to run over medical charting schemas and patient data structures locally.

How I built it: Adapted the Core ML vector search and SQLite FTS5 lexical indexing pipeline from OpenIntelligence. Structured the interface in SwiftUI around a local SwiftData store, integrating SMART on FHIR discovery and OAuth (ASWebAuthenticationSession) to pull demo Patient, Condition, and MedicationRequest resources directly from EHR sandboxes into the local offline model.

Why I built it: Built to migrate off completion-era OpenAI endpoints (following deprecation announcements) to the modern Responses API, adding advanced multi-modal tool support.

How I built it: Programmed a SwiftUI app following the MVVM pattern with AppContainer dependency injection. Created Combine streaming handlers to parse 40+ event types, integrated built-in tools (web search, file search, code interpreter), and implemented security-scoped consent sheets to gate higher-risk actions.

Engineering Trade-Offs: Chose raw SwiftUI MVVM + Combine over heavy third-party state frameworks to minimize binary size. Implemented custom decoders for OpenAI's raw event streams to handle interrupted network state cleanly.

Why I built it: Created because OpenAI's Assistants playground limited assistants to a single vector store, preventing me from chatting with multiple large document databases at once.

How I built it: Created in SwiftUI using async/await. Implemented local PDFKit extraction, Vision OCR document parsing, and metadata chunking. Connected Pinecone's index/namespace endpoints and OpenAI's embeddings, securing sandboxed documents via security-scoped bookmarks and Keychain storage.

Engineering Trade-Offs: Opted for a direct Pinecone REST integration instead of heavy SDKs to preserve memory. Implemented security-scoped bookmarking for local sandboxed files, accepting iOS file-system restrictions while preserving data isolation.

Why I built it: Created to query and digest dense medical device documentation (IFUs) on my iPhone, and to serve as my introduction to Swift development and Apple App Store submission.

How I built it: Programmed in SwiftUI and Combine, mapping assistant endpoints, threads, and runs. Persistent message storage was constructed using local JSON serialization, and the project shipped after resolving ~30 Connect App Review rejections.

Debugging Deep-Dive: App Store rejected initial builds 30+ times due to sandboxing violations when querying local PDF documents. Solved by implementing security-scoped bookmarks (using startAccessingSecurityScopedResource()) and moving document processing to an isolated background thread to prevent UI hangs.