Proof · Problem → System → Outcome

Case Studies

Featured case studies showing production systems with concrete deployment outcomes. Each follows the same structure: context, problem, system design, technical decisions, and measurable results.

Featured Systems

8 case studies

Monthly technical case studies with client testimonials

Each month we publish one deep implementation story with measurable outcomes, architecture notes, and verified testimonial snippets.

Technology: Swift 6Technology: Core MLIndustry: FinanceProject Type: MVP Sprint

March 2026

CalmLedger: Zero-Cloud AI Finance

Core ML categorization, SwiftData sync, and zero cloud API calls.

April 2026

DevScope: Swift 6 Performance Upgrade

200ms analysis loop with actor-based scheduling.

May 2026

Startup MVP: 4-Week Sprint Launch

Shipped on time and supported a $500k seed raise.

Download All Case Studies (PDF)Want similar results? Book a consultation

Apple Platform Market Baseline

The systems below are designed for real Apple-platform constraints and opportunity scale, using only official Apple-published figures.

Apple reports an active installed base of more than 2.2 billion devices. Apple Q1 2024 Results
Apple reports the US App Store ecosystem facilitated $406 billion in developer billings and sales in 2024. Apple Newsroom
Apple reports the App Store prevented more than $9 billion in fraudulent transactions. Apple Newsroom

CalmLedger

FinanceZero Cloud

Privacy-first AI finance app with Core ML categorization, SwiftData sync, and zero cloud API calls

Context

CalmLedger needed transaction categorization, spending insights, and optional multi-device sync for sensitive financial data. The hard requirement: no transaction data could leave the device for inference, analytics, or backend processing.

System Design

•SwiftData local store as the source of truth, with CloudKit private database sync across the user's own devices
•Core ML INT8 transaction classifier compiled into the app bundle for sub-10ms on-device categorization
•Apple Foundation Models generate spending summaries from aggregated category totals, not raw transaction descriptions
•Feature-scoped SwiftUI modules with typed navigation and no shared mutable state across transaction, category, and insight surfaces

Cloud API calls

Categorization

<10ms

Model size

<4MB

Timeline

6 weeks

Outcome

CalmLedger shipped with offline data entry, local AI categorization, local spending insights, and first-review App Store approval with no privacy nutrition label issues. The privacy guarantee is structural: there is no backend path that can receive transaction data.

SwiftUISwiftDataCore MLCloudKit

Deep dive →

Nick

Open Sourcev1.0 Released

Open-source macOS security suite with six detection layers and on-device AI behavioral threat scoring

Context

macOS ships with XProtect, Gatekeeper, and SIP — all signature-based defences that catch known threats after the fact. For developers, sysadmins, and privacy-conscious users this leaves a wide gap: behavioural threats that abuse signed Apple binaries, persistence dropped by compromised installers, and multi-signal attack sequences that only become critical when correlated across monitors.

Problem

Existing macOS security tools either cost $60+/year, require running five or six separate utilities with no unified view, or are enterprise-grade products far beyond what individual users can operate. There was no free, open-source, on-device-AI-powered security suite that replaced all of them in a single native app.

Constraints

—All inference must run on-device — no cloud dependency, no telemetry, no accounts
—Under 1% CPU and 50 MB RAM in steady state — must not affect daily workflows
—AGPL-3.0 licence — every line of code must be publicly auditable on GitHub
—Six independent detection layers must correlate signals without introducing false positives at idle

Solution

•Built six detection layers in a single native SwiftUI app: System Integrity Audit, Persistence Monitor (FSEvents), Network Watchdog, Process Auditor, YARA Scanner, and Camera & Mic Sentinel
•Designed a ThreatCorrelator with a 30-second sliding window across all monitors, feeding a CoreML model for a 0–1 behavioural threat score
•Integrated Foundation Models to generate plain-English alert explanations — no cloud call, no human-written rule translations required
•Released under AGPL-3.0 on GitHub — free forever, fully auditable, community-extensible

Outcome

v1.0 shipped on GitHub under AGPL-3.0. Replaces BlockBlock, LuLu, KnockKnock, and OverSight in a single native app. CoreML inference runs at under 1% CPU in steady state. Foundation Models generates alert explanations without any network call. The only free, open-source macOS security suite with on-device AI behavioural scoring.

The design premise: every alert must be explainable in plain English, with every inference decision auditable down to the model weights.

Technical Highlights

→FSEvents-based Persistence Monitor watching all LaunchAgent and LaunchDaemon directories in real time
→libyara + CoreML heuristics for YARA scanning — no cloud signature updates required
→Network Watchdog: reverse shell and SSH tunnel detection via outbound connection pattern analysis
→Foundation Models on-device LLM for alert explanations — zero tokens leave the Mac

Swift 6CoreMLFoundation ModelsAGPL-3.0

Deep dive →Product page

The Company App

Live

Offline-first operations platform for small business teams managing inventory, orders, and dispatch

Context

Small businesses with 8–25 employees operating across warehouse and office environments. Teams needed a shared operational state for inventory levels, order tracking, and dispatch assignments — without requiring dedicated infrastructure or technical staff to maintain it.

Problem

The team operated across 4–5 disconnected tools: spreadsheets for inventory, messaging apps for team coordination, paper logs for dispatch, and email for client communication. No single source of truth existed for inventory levels or order status. Warehouse and office staff frequently worked from stale data, leading to fulfillment errors, stock discrepancies, and duplicated effort.

Constraints

—Warehouse connectivity is unreliable — the system had to work fully offline and sync when connection restored
—Multi-user concurrent writes required conflict resolution without a custom backend
—Sensitive business data — supplier contacts, pricing, order volumes — could not transit third-party services
—No IT department on the client side — zero-maintenance sync infrastructure was a hard requirement

Solution

•Designed an offline-first data layer using NSPersistentCloudKitContainer — no backend, no server costs, automatic sync through CloudKit
•Structured separate private and shared CloudKit stores to enforce role-based data access at the persistence layer
•Built a unified Core Data schema covering inventory, orders, dispatch assignments, customer contacts, and team tasks in one model
•Implemented iPad split-view layouts optimised for warehouse scanning — large tap targets, minimal navigation depth, offline-aware UI states
•Designed partial sync so field staff receive only their assigned dispatch queue, not the full dataset
•Eliminated all third-party SDKs for data, networking, and sync — every dependency is a first-party Apple framework

Outcome

Deployed across teams of 8–15 employees managing inventory at multiple warehouse locations. Handles offline operation for hours during connectivity outages and syncs automatically on reconnection — no manual intervention required. Consolidated 4–5 operational tools into one, eliminating the data fragmentation that caused fulfillment errors. Stock discrepancies from stale data dropped in the first month of use. No infrastructure to maintain: CloudKit handles sync, Apple handles availability.

The constraint that shaped the entire architecture: warehouse staff cannot wait for a server response. Offline-first was not a feature — it was the design premise.

Technical Highlights

→NSPersistentCloudKitContainer — zero-backend sync through Apple infrastructure
→Offline-first Core Data with automatic merge policy on reconnection
→Separate private/shared CloudKit stores for role-based data isolation
→Partial sync design — field staff receive only their assigned data subset
→Zero third-party dependencies — reduced attack surface and long-term maintenance burden

Built as part of real production work for small teams and independent projects.

SwiftUICore DataCloudKitOffline-First

Deep dive →Product page

SnipToCode

Web Platform + AI

AI-powered design-to-code platform with multi-framework support and SaaS billing

Context

Frontend developers manually implement the same UI patterns repeatedly across projects and frameworks. Design handoffs from Figma add another layer of translation work — the same layout reimplemented by hand for every new stack.

Problem

No tool understood visual design intent and produced idiomatic, framework-specific code without manual annotation. Existing generators produced generic output requiring significant cleanup before it was usable. There was no iterative loop — one shot at generation, no way to refine through conversation.

Constraints

—Output had to be genuinely idiomatic per framework — React hooks differ from Vue composition API; a generic transpiler produces code that looks right but does not follow conventions
—Users expect real-time streaming feedback — a wait-then-reveal model breaks the flow of iterative work
—Credit-based billing required atomic payment processing and per-request usage tracking

Solution

•Integrated Claude Sonnet 4 vision API for design screenshot interpretation — no manual annotation or component labelling required
•Designed per-framework prompt systems generating idiomatic patterns for React, Vue, Angular, SwiftUI, Flutter, and HTML separately
•Implemented real-time streaming output via Server-Sent Events — code appears token-by-token as generation runs
•Built an iterative AI chat layer for post-generation refinement — users describe changes in plain language without re-uploading the design
•Structured credit-based SaaS billing through Paddle with webhook-driven fulfillment and PostgreSQL usage tracking

Outcome

Live SaaS platform with paying customers via Paddle subscription billing. Converts design screenshots into production-ready code across 6 frameworks with real-time streaming output. Reduces design-to-code time from hours to under 2 minutes per component. The iterative chat loop allows full refinement from a single uploaded screenshot — no re-upload needed between iterations.

The hardest part was not the AI integration — it was building prompt systems that produced genuinely idiomatic output per framework rather than surface-level code that looks correct but does not follow each framework's conventions.

Technical Highlights

→Claude Sonnet 4 vision for design interpretation without annotation
→Server-Sent Events for real-time token streaming
→Per-framework idiomatic prompt engineering (6 targets)
→Credit-based SaaS billing with Paddle webhook fulfillment
→PostgreSQL session and per-request usage tracking

Built as part of real production work for small teams and independent projects.

React + TypeScriptClaude Sonnet 4Node.js + ExpressPaddle Payments

View details →

Xcode Doctor

macOS Native

Static analysis tool that diagnoses Xcode project configuration errors before they cause build failures

Context

iOS developers regularly encounter build failures and App Store rejections caused by Xcode project configuration errors — signing mismatches, entitlement gaps, Watch and Widget target dependency issues. These problems are not surfaced by Xcode itself until a build or submission already fails.

Problem

Manual inspection of .xcodeproj XML files, entitlement plists, and signing configurations is slow and error-prone. Developers discovering configuration issues after an App Store rejection face significant delays — Apple review turnaround adds days to each correction cycle.

Constraints

—Read-only by default — the tool must never modify project files, even unintentionally
—Analysis must complete in under 2 seconds to be useful during rapid debug iteration
—Zero telemetry — project configurations contain proprietary app structure and signing information
—Apple notarization required for Gatekeeper-compliant distribution outside the Mac App Store

Solution

•Designed and implemented a static analysis engine that parses .xcodeproj XML directly — no Xcode installation or build process required
•Built 9 specialised checks covering signing certificate validity, entitlement file presence, Watch/Widget target linking, bundle ID consistency, and dependency conflicts
•Structured output with severity categorisation — errors, warnings, and informational findings, each with an actionable description
•Distributed as an Apple-notarized binary with SHA-256 checksum for download integrity verification

Outcome

Runs 9 configuration checks in under 2 seconds. Used by iOS developers to surface signing and entitlement errors before App Store submission — catching issues that previously only appeared after rejection. Debugging time for configuration problems reduced from hours of manual .xcodeproj inspection to a single tool run.

The constraint that drove the architecture: parsing had to be deterministic and zero-side-effect. The tool reads project structure — it never writes, never invokes build commands, never accesses the network.

Technical Highlights

→Static .xcodeproj XML analysis — no Xcode dependency or build invocation
→9 checks completing in under 2 seconds
→Read-only sandboxed access model — write operations architecturally impossible
→Zero telemetry, 100% local processing
→Apple notarization with SHA-256 integrity verification

Built as part of real production work for small teams and independent projects.

SwiftUImacOS NativeXcode AnalysisPrivacy-First

Deep dive →Product page

ECHO Survival AI

Live

Fully offline AI assistant for iOS — local LLM inference with zero cloud dependency

Context

Users who need AI assistance in environments where internet connectivity is unavailable, unreliable, or where sending prompts to a cloud API is not acceptable — field workers, travelers, privacy-conscious professionals, and users in regions with high data costs.

Problem

Every AI assistant app on iOS in 2024 required an active internet connection and transmitted user prompts to cloud infrastructure. There was no production-grade option for private, offline AI interaction on iPhone or iPad. The technical barrier was significant: running a language model locally on a mobile device required solving model storage, memory constraints, inference speed, and battery life simultaneously.

Constraints

—100% offline — no network requests of any kind during inference
—Model storage had to be manageable on devices with 64–256 GB storage without disrupting primary device use
—Battery drain from sustained inference had to be acceptable for practical use — not just a demo
—App Store compliance: Apple's guidelines restrict certain model hosting patterns; distribution required careful review preparation
—Foundation Models framework (iOS 26+) was not available during development — required a cross-version strategy

Solution

•Integrated llama.cpp via Swift bindings — the only production-viable path for local LLM inference on iOS prior to Foundation Models
•Selected GGUF-format quantized models (Q4_K_M and Q5_K_M) — 3–5 GB range that fits on-device without compromising response quality below practical thresholds
•Built battery-aware inference scheduling: background inference pauses when battery drops below threshold; active inference throttles CPU allocation to reduce heat
•Designed model download UX as a first-run flow rather than a gate — users understand the download size before committing, with progress tracking and resume on failure
•Stored models in the app's documents directory using NSFileManager — models survive app updates, do not count against iCloud backup quota
•Zero cloud dependency enforced architecturally — no API keys, no network entitlement required for inference

Outcome

Shipped on the App Store with full offline inference. Users can install a 4 GB model and run open-ended conversations, document summarization, and code explanation entirely on-device. Battery performance on iPhone 15 Pro: sustained inference at 40–60% CPU averages 18–22% battery per hour — acceptable for session-length use. No server costs, no API rate limits, no data transmission.

The technical constraint that defined the architecture: you cannot trade inference quality for model size beyond a threshold — below Q4, users notice degraded output. The solution lives in the quantization-quality curve, not at the extremes.

Technical Highlights

→llama.cpp Swift bindings — local LLM inference without Foundation Models dependency
→GGUF Q4_K_M/Q5_K_M quantization — practical size/quality balance for on-device storage
→Battery-aware inference scheduler — pauses and throttles based on device state
→Resumable model download flow with progress tracking and failure recovery
→Zero network entitlement required at inference time — complete air-gap capability

Built to validate the offline AI architecture pattern before Foundation Models became available.

SwiftUIllama.cppOn-Device AICore ML

Deep dive →Product page

Xcode Localization Translator

Live

AI-powered .xcstrings translation directly inside Xcode — Claude integration within the Source Editor Extension sandbox

Context

iOS and macOS developers who maintain apps in multiple languages. The standard localization workflow requires exporting strings, running them through a translation service externally, and reimporting — context about the app is lost in translation, and the round-trip adds significant friction to iterative development.

Problem

No tool could run AI-powered translation with app context while staying inside Xcode and writing directly to the .xcstrings catalog format. Developers either paid for external localization services (expensive, context-free) or manually ran translations through ChatGPT and copy-pasted results (slow, error-prone).

Constraints

—Xcode Source Editor Extension sandbox: extensions run in a restricted process with limited filesystem access and no direct UI windows
—.xcstrings is a structured JSON format — output integrity was required; partial or malformed JSON corrupts the entire localization catalog
—API key management inside the sandbox: Keychain access from an extension requires App Group entitlements and careful provisioning setup
—The extension must not block the Xcode UI thread — translation requests had to be async with visible progress

Solution

•Built as a proper Xcode Source Editor Extension — invoked from Xcode's Editor menu, operates on the active .xcstrings file
•Claude API (claude-3-5-haiku) for translation — structured output prompting ensures the model returns only the translated value, not surrounding explanation
•Custom .xcstrings JSON parser and generator — reads the existing catalog, inserts translated entries for each target language, writes back atomically
•Keychain API key storage via a shared App Group — the main app handles key entry via a settings panel; the extension reads from the shared Keychain without requiring UI
•Translation runs on a background queue; Xcode editor buffer is updated only when the full response is validated against the expected schema
•Supports batch translation (all missing keys across all target languages in one invocation) and single-key translation with language selection

Outcome

Reduces localization time from hours of copy-paste work to minutes of in-editor batch translation. The structured output approach means translated strings match the .xcstrings format exactly — no manual formatting, no catalog corruption. Developers keep full context (app name, feature area, adjacent strings) in the prompt, which meaningfully improves translation quality for context-dependent strings like button labels and error messages.

The hardest part was not the AI integration — it was learning the exact boundary of what a Source Editor Extension is permitted to do. Keychain access from a sandboxed extension requires getting the App Group entitlements exactly right, and a single misconfiguration silently fails without an error message.

Technical Highlights

→Xcode Source Editor Extension — runs inside Xcode, modifies active file directly
→Claude API with structured output prompting — translation-only responses, no prose
→Custom .xcstrings JSON parser/generator with atomic write-back
→App Group Keychain sharing — secure API key access from sandboxed extension
→Batch and single-key translation modes with background async execution

Demonstrates how to integrate AI capabilities within Xcode's Source Editor Extension sandbox constraints.

Xcode ExtensionClaude API.xcstringsSwift

View details →

Filmforge

New

Native macOS video toolkit — 18+ tools built on AVFoundation with hardware-accelerated export, free forever

Context

Mac users who need to perform common video operations — resize for upload, compress for email, trim a clip, strip audio, replace a soundtrack — without learning FFmpeg syntax or installing Electron-based apps that ship with ads, watermarks, or subscription paywalls.

Problem

macOS ships the best media frameworks in the industry: AVFoundation, VideoToolbox, CoreImage. But there is nothing between QuickTime (trim only) and iMovie (full non-linear editor) for everyday utility tasks. Every attempt to fill the gap was FFmpeg wrapped in a dated UI, a cross-platform Electron port with no hardware acceleration, or a subscription app that watermarks exports until you pay.

Constraints

—No third-party media libraries — App Store sandbox and GPL licensing made FFmpeg incompatible; entire stack had to be first-party Apple frameworks
—18+ tools had to compose without repeated re-encoding — each additional operation must not degrade quality
—Real-time CIFilter preview during playback — GPU-accelerated filters had to run at full frame rate on M1 or later
—Live file size estimation before export — users needed projected output size before committing to a long encode

Solution

•Non-destructive operation stack — Resize, Trim, Compress, Blur, Replace Audio stored as metadata until export; source file never modified
•Single-pass export via AVAssetWriter — entire stack composes into one encoding pass, eliminating generation loss from repeated re-encoding
•VideoToolbox hardware H.264/H.265 encoding — Apple Silicon dedicated media engines handle the encode, not CPU cores
•Passthrough mode for audio-only operations — no video frames decoded or re-encoded, finishing in seconds regardless of video length
•CoreImage GPU filter pipeline for real-time preview — brightness, contrast, blur, sharpen at full playback frame rate without export

Outcome

18+ video tools in one native macOS app, free on the Mac App Store. Hardware-accelerated H.265 export completes a 1-minute 4K clip in approximately 8 seconds on M2. Audio-only operations finish in seconds on any clip length. The operation stack composes cleanly — Resize + Brightness + Trim + Compress with no quality degradation. Zero network access, no analytics, no watermarks, no subscription.

The key architectural decision: treat the operation stack as metadata, not a progressive pipeline. Defer all processing to a single AVAssetWriter composition pass. This eliminated the re-encoding quality loss that makes video utility tools feel cheap.

Technical Highlights

→Non-destructive operation stack — metadata until single AVAssetWriter export pass
→VideoToolbox H.264/H.265 hardware encoding via Apple Silicon media engines
→CoreImage GPU filter pipeline for real-time preview at full playback frame rate
→Audio passthrough for Remove/Extract — no video re-encode, seconds to complete
→Pure SwiftUI + SwiftData — no AppKit wrappers, no web views, zero cloud dependency

Built as proof that native, privacy-first utility software can be both powerful and free.

SwiftUIAVFoundationVideoToolboxCoreImageSwiftData

Deep dive →Product page

More Projects (6 additional)

HobbyIt

iOS + AI

Health and habit tracker with Apple Intelligence integration. HealthKit sync with on-device AI suggestions.

View details →

KetoDietPro

iOS

Ketogenic diet tracker with macro calculation engine. Offline-first Core Data with barcode scanning.

View details →

SwiftUI CrossPreview

Xcode Extension

macOS/iOS cross-preview extension for SwiftUI. Simultaneous platform previews in Xcode.

View details →

SwiftUI Templates

iOS Components

Production-ready SwiftUI component library. Covers authentication, navigation, forms, data display.

View details →

AI-native OS Experiments

Research

Shell automation with local LLMs. Ollama integration for privacy-first command generation.

View details →

Need a System Like This?

These case studies demonstrate the approach to production-grade systems. Clear architecture, privacy-first design, and outcomes that matter.

New engagements start with a structured application to ensure strong product and technical alignment.

Start a Project View service offerings

Last updated: April 6, 2026