Get it to run at all
This was copied API references, Playground threads, red Xcode errors, and rebuilding until the app finally stopped breaking.
OnSite Specialist · 4 App Store Apps · Offline Apple Intelligence
Gunnar Hostetler
I have 4 iOS apps on the App Store because it's fun. I build solutions to my problems and find great joy if it helps others. OpenIntelligence is the one I am most proud of: an Apple Intelligence neural RAG engine that runs completely offline, airplane mode included, with three quality modes: Standard, Deep Think, and Maximum.
About Me
In a very, very small nutshell, I build solutions to my problems and find great joy if it helps others.
At work, I provide intraoperative technical support to Stanford surgeons at the VA in Palo Alto as an OnSite Specialist for Stryker. There is no Stryker management on site. It is a full-service contract, which means I handle endoscopy towers, cameras, instruments, cleaning, wrapping, and doing it all over again.
On my own time, I have a hand in almost every jar of relevant code languages with a larger focus on Swift/iOS because it's always on you and easy to test in the real world. I currently have four live on the App Store, all built entirely in Swift and SwiftUI.
ChatGPT-4 blew my mind when it first showed up because it felt like a gap filler for understanding hard things fast. I was already working with a lot of IFUs and medical documentation at Stryker, got ultra curious about how these devices worked, and started using AI constantly to learn faster.
Most of my early build process was primitive. I copied API docs into text files, fed them into the Playground, pasted red Xcode errors right back into the loop, rebuilt, and kept doing that until the apps finally ran. AppStoreConnect is like a mini-game at this point. My first app was rejected 30 times, my second 10 times, and my third made it through App Review on the first pass because of how many precautions I took before even attempting.
Before this job, I located utilities and had zero hospital experience. Now I have hospital experience, and I can see exactly what could be built.
How The Apps Changed
OpenAssistant
OpenAssistant started because the official ChatGPT iOS workflow was helping me learn dense docs fast, but five files at a time was not enough and the whole thing started feeling slow and cramped. I found the Assistants API, saw people building wrappers around it, and thought: why would I pay for somebody else's app if I could make my own? Then it was copied docs, Playground threads, red Xcode errors, rebuilds, and around 30 App Review rejections before it finally made it onto the App Store.
OpenCone
OpenCone happened once the Assistants Playground started feeling cramped again. Bigger queries, larger document sets, and one vector store per assistant kept making me think there had to be an easier way. I found Pinecone, learned indexes, namespaces, and embeddings on the fly, and built it because I wanted one thread to reach across more material with more control. Getting it onto the App Store was also me wanting to prove I could do it again.
OpenResponses
OpenResponses happened when I realized the old completions-era setup was getting deprecated and my first app was going to become useless if I left it there. I had OpenAssistant open in one VS Code window, OpenResponses in another, and rebuilt the core flow piece by piece on Responses until it was current again. It was the first one that made it through App Review on the first submission.
OpenIntelligence
OpenIntelligence is the app I am most proud of. I wanted to see how far I could push the same document workflow on Apple's local model path, so I built an Apple Intelligence neural RAG engine that runs completely offline. The big constraint was the 4,096-token session budget, so I went deep on the docs and built a recursive multi-session reasoning loop into the app.
What Kept Happening
Something starts annoying me, I go way too deep on the docs, and I keep building until it does what I wanted in the first place.
Get past one vector store
The OpenAI Playground started feeling cramped, so I went looking for more retrieval control and ended up knee-deep in Pinecone indexes, namespaces, and embeddings.
Keep the first app alive
Once the old endpoints were on the way out, this became a migration project. Old app in one window, new one in another, and a lot of nudging to keep the logic from drifting.
Take it fully offline
I wanted to take the same document workflow onto Apple's local model path, deal with the tighter limits, and make it work completely offline.
Skills
Most of this came from work, documentation, repetition, and building things until they worked. The hospital side made reliability matter. The app side gave me a place to keep testing ideas.
Patterns
Documentation
Debugging
Iteration
Workflow Translation
Retrieval Design
Release / Review
RAG Pipelines
API Integration
Privacy Boundaries
How I Build
Swift / SwiftUI
Apple Intelligence
Foundation Models
OpenAI APIs
Python
Pinecone
Qdrant
Xcode
App Store Connect
Git / GitHub
VS Code + Copilot
Documentation Ingestion
Healthcare Ops
Intraoperative Support
HIPAA Compliance
Stakeholder Coordination
Medical Device Support
Procurement
Instrument Tracking
Surgical Workflows
Selected Work
These are the main apps. Together they show the same loop repeating: hit a limit, get annoyed, read more docs, and build around it.
Why These Four
Each app came from a specific limit
It was basically the same pattern every time: hit a ceiling, get curious, and build around it instead of accepting it.
The repos and App Store pages are public if you want the full trail.
OpenAssistant
ChatGPT's iOS workflow, but with more control.
OpenCone
One thread, more documents, and more retrieval control.
OpenResponses
Move the whole thing to Responses before the old stack died.
OpenIntelligence
Take the same document workflow offline on Apple's local model path.
OpenResponses
OpenResponses started when I realized the old completions-era setup was getting deprecated and my first app was going to become useless if I left it there.
- Why it exists: I was not going to let the first app die on old endpoints.
- How I built it: I had OpenAssistant open in one VS Code window, OpenResponses in another, and started swapping the core over to Responses piece by piece.
- What changed: Responses gave me a much bigger playground: better tool calls, file search, web search, streaming, reasoning, and more room to keep iterating.
- Why it matters: It kept the app current and was the first one that passed App Review on the first submission.
SwiftUI
OpenAI Responses API
Request Inspection
Approval Flows
File Search
Web Search
Code Interpreter
Async/Await
OpenIntelligence
OpenIntelligence is the app I am most proud of. I wanted to see how far I could push the same document workflow on Apple's local model path, so I built an Apple Intelligence neural RAG engine that runs completely offline.
- Why it exists: I wanted the same grounded document workflow, but local, private, and usable in airplane mode.
- How it works: It imports PDFs, images, and code, builds a local knowledge base, and answers with source-backed responses on-device. It has three quality modes: Standard, Deep Think, and Maximum.
- What got hard: Apple's public on-device sessions are capped at 4,096 tokens, and that budget has to cover instructions, retrieved context, tool and schema overhead, and the answer itself. I did a ton of research on how to work around that ceiling and built a recursive multi-session reasoning loop into the app.
- Why it matters: It is the app that made me think, "I could do this on a much larger scale."
Apple Intelligence
FoundationModels
SwiftUI
Swift 6
On-Device RAG
Vision OCR
Async/Await
OpenCone
OpenCone came from hitting the OpenAI Playground ceiling again. Bigger queries, larger document sets, and one vector store per assistant kept making the whole thing feel cramped.
- Why it exists: I wanted one thread to reach across more material with more retrieval control than the Assistants setup was giving me.
- How I built it: It was basically the same process as the first app, except now the docs pile included Pinecone and I was learning indexes, namespaces, and embeddings as I went.
- What changed: One assistant being tied to one vector store stopped feeling acceptable. I wanted more room and more control.
- Why it matters: Getting it onto the App Store was personal proof that I could keep pushing the retrieval side further.
SwiftUI
OpenAI Embeddings
Pinecone API
PDFKit
Vision OCR
RAG
Async/Await
OpenAssistant
OpenAssistant started because the official ChatGPT iOS app made dense docs way easier to work through, but five files at a time was not enough and the workflow hit its ceiling fast.
- Why it exists: I was newly at Stryker, working through piles of IFUs and medical documentation, and got obsessed with learning that world faster.
- How I built it: I found the Assistants API, saw people building wrappers around it, and thought, why would I pay for somebody else's app if I could make my own? Then it was copied API docs, Playground threads, red Xcode errors, rebuilds, and repeating that loop until the app ran.
- What got in the way: Around 30 App Review rejections forced me to learn review, polish, and edge-case cleanup while I was still learning iOS itself.
- Why it matters: It is still the foundation for everything that came after it.
SwiftUI
OpenAI Assistants API
Vector Stores
Code Interpreter
File Search
Professional Experience
Independent iOS Projects
Personal apps and prototypes
Nov 2023 – Present- Shipped Work: Four live App Store apps, all built entirely in Swift and SwiftUI.
- Build Style: Copied API docs into text files, used the Playground constantly, pasted red Xcode errors back into the loop, rebuilt, repeated.
- App Review: AppStoreConnect is like a mini-game at this point. First app rejected 30 times, second 10 times, third passed App Review on the first submission.
- Signature Build: OpenIntelligence is the one I am most proud of: offline Apple Intelligence, three quality modes, and a recursive multi-session reasoning loop I built after a lot of research around the 4,096-token limit.
- Traction: All four apps combined have over 1,000 downloads, and OpenIntelligence has generated $190 since its February 2026 release.
I have lots to learn, and I genuinely want to get better at all of this.
OnSite Specialist
Stryker | VA Palo Alto Health Care System
Aug 2022 – PresentI provide intraoperative technical support to Stanford surgeons at the VA in Palo Alto as an OnSite Specialist for Stryker. There is no Stryker management on site, so I am the first and only line of defense that is not in an office across the city.
- Full-Service Contract: I handle endoscopy towers, cameras, instruments, cleaning, wrapping, and doing it all over again.
- Operational Range: I understand HIPAA, surgical workflows, procurement, stakeholder relationships, par levels, and repair and replacement contracts.
- Daily Reality: Every day is different. New tech comes in, new residents come in, new fellows come in, and all of them have specific requests for video setups and instrumentation.
- Coverage: I support 8 operating rooms and work directly with surgeons, OR staff, nurse managers, SPS, BioMed, residents, fellows, and hospital stakeholders.
I had zero hospital experience before this job. Now I have it, and I can see exactly what could be built.
Data Analyst
Above and Beyond Family Recovery Center
Jan 2020 – Jan 2021Managed and validated patient-facing operational data in a clinical environment where accuracy, confidentiality, and process reliability mattered.
- Data Quality: Managed and validated data files for roughly 100 patients per week, ensuring accuracy and compliance.
- Process Improvement: Streamlined data handling workflows to improve operational efficiency and reliability.
- Confidentiality: Maintained strict care while handling sensitive diagnoses, insurance data, and internal records.
Utilities Locator
Utility Resource Group, LLC
Dec 2016 – Dec 2020Field operations role built around precision, safety, route planning, and independent execution in a high-accountability environment.
- Safety & Precision: Located natural gas lines with specialized equipment while maintaining safety compliance.
- Operational Throughput: Navigated mapping systems and company tools to manage daily work, averaging 24 tickets closed per day.
- Training: Mentored and trained 25 new technicians on safety protocols and field best practices.
Physical Therapy Aide
Athletico
Sep 2015 – Oct 2016Supported patient care workflows while handling scheduling, insurance administration, and front-line clinic operations.
- Patient Support: Assisted physical therapists with patient care and rehabilitation exercises.
- Scheduling: Managed appointment changes, cancellations, and day-to-day clinic coordination.
- Insurance Workflow: Verified, scanned, and processed insurance information across Medicare, worker's compensation, and private insurers.
District Manager
Aqua-Guard Management Inc
Mar 2010 – Sep 2015Early leadership role managing distributed operations, staffing coverage, and on-call execution across multiple sites.
- People Management: Supervised seven managers and 31 lifeguards across nine pools.
- Coverage & Reliability: Ensured all shifts were covered, stepping in directly when required.
- Operations: Built daily routes and managed pool maintenance through chemical administration and round-the-clock support.
Education
Western Illinois University
Aug 2010 – Dec 2014Bachelor's Degree in Kinesiology and Exercise Science.
Get In Touch
I want to build where this matters
I want a space where I can focus this energy on building solutions that impact thousands.
If you are in healthtech, medical devices, or on-device AI, or if you are building something where clinical background directly multiplies technical output, I want to talk.