I develop my own products and build software for clients.

What I built

• Local data storage and full-text search through a local-first architecture with on-device storage and indexing.
• Orchestration of legal research and document analysis through an agentic tool loop with bounded sub-agents and parallel searches.
• Processing of heterogeneous file formats through an extensible processor pipeline for formats including PDFs, Office documents, emails, and email archives, combining local extraction with cloud-based vision models.
• Semantic search and AI-powered analysis across deterministically selected document sets, consolidating results into structured comparison tables.
• Hallucination detection and prevention through a structured citation protocol, programmatic validation against original documents, and dynamic context injection between agent-loop iterations.
• Indexing of statutes and case law through a dedicated cloud indexing service with a research corpus kept separate from client data.
• Regional restriction of model processing by routing all model requests to endpoints in EU regions.
• Integration of text and spreadsheet editors built on a shared Lexical document model, with change tracking for AI-generated edits and export pipelines for DOCX, PDF, Markdown, XLSX, and CSV.

What I built

• End-to-end type safety between the iOS app and Vapor backend through shared Swift models and a typed API router.
• Separation of the user interface from business logic and platform services through modular TCA features, UIKitNavigation, and injectable client interfaces.
• Synchronization of locally recorded training data through a SQLite outbox, server-side sequence numbers, deletion tombstones, and timestamp-based conflict resolution.
• Validation of AI-generated training plans through structured model output, deterministic invariant checks, and a correction attempt with specific error feedback.
• Streaming of contextual AI responses through server-side preparation of training history and an SSE pipeline between the model provider, Vapor, and the iOS app.
• Verification of reducers, API contracts, and model communication through Swift Testing suites, TCA TestStore, and local HTTP test servers for streaming responses.

What I built

• Spoken training dialogues through a cascaded pipeline combining audio capture, Azure speech recognition, AI response generation and OpenAI speech synthesis.
• Speech input segmentation through custom voice activity detection with an adaptive noise threshold, audio buffering and pause detection.
• Configurable role-play through separate models for scenarios, personas, difficulty levels and conversation goals.
• Feedback on individual responses through a separate AI call with conversation context, configurable assessment criteria and structured JSON output.
• Updates to client-specific training content through JSON imports from a web service, QR-code entry and local scenario storage.
• Coordination of recording, transcription, response generation and playback through asynchronous event streams and explicit conversation states.

What I built

• Detection of different diagram elements through an RF-DETR detector trainable on annotated COCO data, followed by conversion into structured nodes.
• Reconstruction of connections through line-mask skeletonization, arrow endpoint extraction, and geometric association with nearby shapes.
• Assignment of rotated labels through PaddleOCR, orientation detection, and spatial matching of text regions to detected nodes.
• Estimation of shape colors and orientation through SAM2 segmentation, masked color analysis, and specialized ResNet18 angle classifiers.
• Serving inference through a containerized FastAPI endpoint with locally supplied model weights and structured JSON output.
• Reproducible model and pipeline development through training notebooks, nbdev module export, and Terraform-managed SageMaker infrastructure.

What I built

• Modernization of the weather feed and detail views through modular SwiftUI and TCA components with explicit state logic and snapshot tests.
• Restructuring of location management and search through separate services for storage, location resolution, and search suggestions, alongside server-side management of severe weather notification subscriptions.
• Integration of configurable widgets using location-specific weather data and deeplinks with correct navigation even on a cold app launch.
• Accessibility improvements through adjustable font sizes, contrast-aware rendering, and alternative controls for horizontal scrolling.
• Restructuring of advertising and consent integration through separate services, explicit loading states, and on-demand ad loading in the weather feed.
• Modularization of the Apple Watch app using SwiftUI and TCA with injectable services, plus reliable location resolution between Watch and iPhone through fallbacks, timeouts, and deterministic tests.
• Integration of Rust-based radar and SkyScene rendering through Swift bindings and Metal, with rendering controlled by visibility and the app lifecycle.
• Reproducible GPU analysis through automated UI flows, local radar fixtures, and comparative Metal System Traces on physical devices.
• Automation of build, test, and release processes using GitLab CI and Fastlane with reusable build artifacts, unit, snapshot, and UI tests, and aggregated coverage reports.

What I built

• Detection of names and other entities through token classification running locally in the browser with Transformers.js.
• Supplementary pattern matching for email addresses, phone numbers, URLs and numbers using regular expressions.
• Restoration of original details from placeholders through a mapping held in the browser.
• Integration into existing websites through a JavaScript module with its own interface and style isolation using Shadow DOM.

What I built

• Communication with proprietary door controllers through a BLE channel with sequential request processing, response parsers, and checksums.
• Decoupling hardware access from the interface through protocol-based services, dependency injection, and reactive state propagation with Combine and RxSwift.
• Coordination of firmware installation through sequential transfers, a progress model, and explicit cancellation and error handling.
• Integration of video support through Amazon Chime with API-based session coordination and retries for invalid meeting configurations.
• Local persistence of support conversations through a Core Data model for tickets, messages, and attachments with separate API and file services.
• Verification of protocol parsers, disconnect handling, and call coordination through XCTest and generated service mocks.

What I built

• App modularization through separate Swift package targets for features, data access, and the design system, with composed TCA reducers.
• Local muscle catalog through Realm models with relationships between muscle groups, muscles, and exercises, plus initial data seeding.
• Asynchronous data access through an actor-bound Realm client and injectable dependencies for feature reducers.
• Integration of purchase and entitlement queries through a RevenueCat adapter that bridges callback APIs to async/await.

What I built

• Modeling complex interval sequences through separate timer, state, and session models with Combine publishers for pauses, rounds, and interval transitions.
• Local management of workouts and intervals through Core Data with a separate data access layer and injectable services.
• Synchronization of shared workouts over SharePlay through state messages for timer mode, elapsed time, and interval position, with state transfer for new participants.
• Coordination of HealthKit recording with workout progression through an actor that handles state changes, pauses, and interval transitions.
• Verification of timer states, data access, and purchase entitlements through behavior-based unit tests, mocked services, and local StoreKit tests.

What I built

• App structure through separate Swift package targets for search, watchlists, detail views, and data access, with TCA state models.
• Modeling nested company and balance-sheet data through typed Codable structures for company information and quarterly and annual figures.
• Bilingual interfaces through a dedicated localization client with German and English resources bundled in the package.

What I built

• Cross-platform implementation of the interface through a shared codebase using React and Ionic.
• Integration of the hardware product with the companion app over Bluetooth Low Energy.