Draft: highlighted text is inferred reasoning that still needs confirming
MUKESH SHARMA
Open to senior and lead iOS roles · Canada, US (TN), remote

I build iOS apps that keep working when conditions don't.

AI answers that render while they're still being written. Payments that go through with no signal. Clinical AI that never sends a patient's voice off the phone. I've been shipping iOS since 2011 across banking, retail, payments, health and smart cities. Below are the decisions behind that code, not just the job titles.

15
years shipping iOS, since 2011
30+
production apps shipped
~20
apps built on a framework I designed
≈0
offline payment failures after the queue
↘ Selected work

Five problems, and how I reasoned through them.

Each card opens the full story: the constraints, the options I turned down, the design I chose, and what I'd change today.

01 / Fintech AI assistant

Answers that appear as they're written, without a single data race.

I built the real-time LLM streaming layer and the chat surface for a generative-AI investing assistant inside a flagship iOS app: eight Swift/SwiftUI modules, from the entry point to the tokenizer.

8 modules8 response typesiOS 17 → 26
Swift 6 · SwiftUI · Combine · actors · Alamofire streaming · Swift Testing
02 / TruPOS payments

Field sales kept taking payments with no signal.

I led the 0→1 SwiftUI checkout hub, built around an offline transaction queue that syncs when the device reconnects.

72h offline retention3 apps, 1 package20–50% faster releases
SwiftUI · Core Data · BGTaskScheduler · Bluetooth/MFi · Fastlane
03 / Dr Godly

Clinical AI that never sends a patient's voice to a server.

Apple Foundation Models fine-tuned with LoRA for clinical vocabulary, plus live two-party transcription in video consults. Built solo across iOS, Android and the backend.

On-device inference6+ repos, soloProduction pilot
Foundation Models · LoRA · Apple Speech · Twilio Video · Node.js
04 / NEOM Smart City

One framework under ~20 apps, so no team rebuilt networking or auth.

I designed the core Swift package for networking, OAuth, SSL pinning and shared UI, and shipped the Smart City app's first App Store release.

~20 adopting appsFirst App Store release3 engineers mentored
SPM · Apollo GraphQL · OAuth 2.0 · SSL pinning
05 / Kroger

The release pipeline the whole iOS org shipped with.

I wrote the build, sign and deploy pipelines for four grocery banner apps, and shipped the Gamification module to the full customer base behind feature flags.

4 banner appsOrg-wide adoptionv1.0 → v2.0
Fastlane · GitHub Actions · Ruby · Remote Config · Crashlytics
Full history: 11 roles, 2011 → 2026
  • Mar 2026 – Aug 2026
    Senior iOS Engineer · Publicis Sapient
    Generative-AI investing assistant for a wealth-management client (case study 01).
  • Dec 2024 – Mar 2026
    Lead iOS Engineer · Tech Mahindra
    TruPOS enterprise payments app (case study 02).
  • Dec 2024 – Mar 2026
    Senior Mobile Engineer, React Native · Tech Mahindra
    TruGreen consumer app on iOS and Android, with native Swift modules for biometrics, deep links and secure storage.
  • Jan 2024 – present
    Founder & Lead Engineer · Dr Godly
    On-device AI for clinical consultations (case study 03).
  • Jun 2023 – Jan 2024
    Senior / Lead iOS Engineer · NEOM via Tech Mahindra
    Smart City app and the core framework (case study 04).
  • Jan 2023 – Jun 2023
    Senior / Lead iOS & Release Engineer · Kroger via Tech Mahindra
    Gamification module and org-wide release pipelines (case study 05).
  • Jun 2022 – Dec 2022
    Senior iOS Engineer · Gap Inc. via Avenue Code
    Features for Gap, Athleta, Banana Republic and Old Navy; led the Notification System workstream from scratch to launch.
  • Sep 2021 – Jun 2022
    Senior iOS Engineer · TD Bank via Clearbridge Mobile / Amdocs
    Three core banking modules, Face ID / Touch ID sign-in, jailbreak and tamper hardening, and contributions to TD's iOS Banking Foundation Framework.
  • Nov 2017 – Aug 2021
    Lead iOS Engineer · Tata Consultancy Services
    Trickl smart-utility app; Green Button apps for Toronto, London and Whitby Hydro.
  • Jan 2015 – Nov 2017
    iOS Engineer · Tata Consultancy Services
    Six internal Pfizer apps with custom Core Graphics charts, Janrain SSO and offline sync, including multi-caregiver conflict resolution in GroAssist.
  • Aug 2011 – Dec 2014
    iOS Engineer · GlobalLogic, smartData, Regard Solutions
    8+ App Store apps across healthcare, retail and consumer, plus accessibility R&D for Qualcomm Healthy Circle.

Toolbox. Swift 6 and 5.x, Objective-C, SwiftUI, UIKit, Combine, async/await, actors, Sendable, Core Data, Core Animation, Core Graphics, App Intents, WidgetKit, TipKit · MVVM-C, Clean/VIP, protocol-oriented design, SPM modularization · Apple Foundation Models, Core ML, LoRA, LLM streaming · URLSession, Alamofire, GraphQL (Apollo), OAuth 2.0, Keychain, SSL pinning · XCTest, Swift Testing, XCUITest, Instruments · Fastlane, GitHub Actions, Xcode Cloud, Crashlytics, Sentry · React Native, Node.js/TypeScript.

Education. B.E., Computer Technology, Priyadarshini College of Engineering, R.T.M. Nagpur University, 2011.

↘ Products

Things I've built, and am building.

Solo and end to end: the client, the backend and the release. These are where I test ideas before I'd recommend them to a team.

Personal finance · iPhone

Penny TestFlight beta

A money coach you can just ask. Link your bank once, then ask about your spending in plain words, by typing or by voice, and get the answer back as a sentence and a chart.

Swift 6 · SwiftUI · Swift Charts · App Intents · Live Activities · AWS · Plaid · Claude
Sales automation · macOS + Linux

CarSales Bot

A WhatsApp sales agent for car dealerships. It answers inventory questions around the clock through an OpenAI tool-calling agent and hands serious buyers to a human. It ships as a signed, notarized desktop app so dealers can install it themselves.

TypeScript · Electron · Baileys · SQLite · 290+ Vitest tests
Clinical AI · iOS + Android

Dr Godly Pilot

On-device AI for doctor–patient consultations: live two-party transcription and note summaries from a LoRA-tuned Apple Foundation Model. Patient audio stays on the phone.

Foundation Models · Apple Speech · Twilio Video · Node.js
Testing the Penny beta, and what happens to your data

Download and test it

The beta runs on Penny's test server with a test bank. No real bank account is needed, and none should be used.

  1. On an iPhone running iOS 26 or later, install TestFlight.
  2. Open testflight.apple.com/join/zxSFk3yd on that iPhone, tap Accept, then Install. The app appears as Penny Dev.
  3. Tap Sign in or create account and create an account with your email. You'll get a code by email to confirm it.
  4. Set up Face ID, then link a bank. Pick any bank in Plaid's window and use Plaid's public sandbox login:
    Usernameuser_goodPasswordpass_goodCode, if asked1234
    These only work in the test environment and connect to sample data.
  5. Ask: "How much did I spend each day this week?", "Break down my spending by category" or "Which subscriptions am I paying for?"
  6. Optional: say "Hey Siri, talk to Penny Dev" to ask by voice.
  7. Send feedback from TestFlight → Penny Dev → Send Beta Feedback.

Your data and privacy, plainly

  • Kept in Canada. Everything Penny stores lives in its own AWS account in the Canada (Central) region, encrypted in transit and at rest.
  • No ads, no tracking, no selling. No advertising, analytics or tracking code.
  • Your bank password stays with Plaid. The link is read-only; Penny can't move money.
  • Short-lived chat. Encrypted on the phone, never backed up, and cleared after 30 minutes, on a new chat, or on sign-out. Server-side context is deleted after 30 minutes too.
  • How answers are written. Your question and the relevant figures go to Anthropic's Claude, processed in the United States. Card numbers, PINs and passwords typed into chat are blocked.
  • Retention. Account data is kept while the account is open; AI request logs for 1 year and server logs for 90 days. You can ask for deletion at any time.

Full details are in the MeetPenny app privacy policy. Penny explains your own spending. It isn't financial, tax or legal advice, and it's for people 18 and older.

↘ How I work

Principles, with receipts.

Four habits that show up in every case study above. Each one links to where it actually happened.

Agree on the contract before the code.

I co-designed the streaming schema with backend and AI-platform teams before writing a parser, and gave a ~20-app framework explicit interface contracts and a versioning strategy.

Design for the bad day.

No signal at a customer site. A response type the app has never seen. A stream arriving faster than the UI can draw. Those cases set the architecture, not the happy path.

Build the piece other teams reuse.

A core package under ~20 apps, a payments package that ended rewrites in 3, and a coach-mark framework other engineers adopted. Leverage beats heroics.

Make releasing boring.

One pipeline for four banner apps, release cycles 20–50% shorter, and risky features behind flags and staged rollouts. Shipping should be the calm part.

↘ Contact

Shipping something on iOS that has to just work?

I'm looking for senior and lead iOS roles on product teams doing serious work with Swift, on-device ML or AI on iOS. I'm based in Toronto and London, Ontario, and I'm a Canadian citizen. I can relocate immediately to San Francisco, New York or Seattle, and I'm TN-eligible for US roles, with no lottery or petition needed.

01 / Fintech AI assistant

Answers that appear as they're written, without a single data race.

Senior iOS Engineer (contract) · Publicis Sapient · Mar 2026 – Aug 2026 · Toronto, hybrid

A wealth-management firm was adding a generative-AI investing assistant to its flagship iOS app. Answers had to stream in token by token, like a consumer chat app, but they also had to render tables, lists and charts, in a financial app where a half-drawn table is worse than none.

Constraints

Options I weighed

The design

  1. Alamofire DataStreamRequestRaw bytes from the backend-for-frontend, as they arrive.
  2. actor JSONLTokenizerAccumulates a buffer, splits on newlines, extracts only complete objects.actor isolation: no shared mutable buffer
  3. Decode on a component discriminatorEight types, validated at the parse boundary (table header/row parity, chart value bounds). Unknown types degrade gracefully instead of crashing.
  4. Orchestration service · custom @globalActorSession lifecycle, chunk-by-chunk message assembly, handoff from loading state to first chunk, cancel and retry.
  5. ChatViewModel · @MainActorUI state is only touched on the main actor; everything crossing the boundary is Sendable.
  6. SwiftUI chat package50+ files covering messages, markdown, tables, lists, charts, typing and retry. Later reused for live-agent chat.
Wire valueRenders as
chat-responseMarkdown text bubble
tablePlain table grid
ordered-list / unordered-listNumbered or bullet list
line-chart / bar-chart / pie-chartChart views with their own ViewModels
suggested-user-intentsSuggestion pills

The data races were the real problem. The first version processed chunks off a callback queue, and races surfaced under load. Moving the tokenizer into an actor, serializing stream processing on a custom global actor, enforcing @MainActor at the UI boundary and making every crossing type Sendable eliminated them.

Outcome

Phase 1Shipped to production on a staged rollout
0Data races under load after the actor redesign
17 → 26iOS range the coach-mark framework supports

What I'd do differently

Put an explicit schema version in the stream from day one. Graceful degradation keeps the app from breaking, but it also hides when the client has fallen behind the backend.

Swift 6 · SwiftUI · Combine · async/await · actors · @globalActor · Sendable · Alamofire · SPM · XCTest · Swift Testing
02 / TruPOS payments

Field sales kept taking payments with no signal.

Lead iOS Engineer · Tech Mahindra · Dec 2024 – Mar 2026

The client needed one iOS app to be the checkout hub for several business lines. Its users are field sales staff taking card payments at customer sites, often in basements, rural properties and other places with weak or no signal. Every payment that failed there was lost revenue and an awkward moment with a customer.

Constraints

Options I weighed

The design

  1. Checkout · SwiftUI + MVVMDynamic, form-based screens for checkout, history and merchant onboarding.
  2. PaymentServiceBluetooth card reader with background modes and a retry layer for flaky accessory connections.
  3. OnlineSubmit to API, then poll status live (Combine + URLSession) and update the UI as it settles.
    OfflineWrite to the Core Data queue: status, attempt count, created-at (≤ 72h).
  4. BGTaskScheduler + reconnect triggerDrains the queue with retries and marks each payment synced.client-generated transaction ID, so a retry can't charge twice
  5. Shared Swift packagePayments, receipts and deep linking, used by three internal apps.

Outcome

≈0Offline transaction failures
3 appsOn one payments package, with no more per-app rewrites
20–50%Shorter release cycles after automation

What I'd do differently

Show the queue to the user sooner: something like "3 payments waiting to sync" on the home screen. Trust in an offline system comes from seeing it work.

Swift · SwiftUI · Combine · async/await · Core Data · BGTaskScheduler · Bluetooth / MFi · App Intents · Fastlane · GitHub Actions · Xcode Cloud · Instruments
03 / Dr Godly

Clinical AI that never sends a patient's voice to a server.

Founder & Lead Engineer · Jan 2024 – present · limited production pilot with early customers

Doctors lose consultation time to note-writing. An assistant that transcribes and summarizes helps, but in healthcare, sending patient audio to a third-party model creates privacy and trust problems. I wanted to know whether the whole loop could stay on the phone.

Constraints

Options I weighed

The design

  1. Twilio Video consultationDoctor and patient on a secure call.
  2. Apple Speech · doctorReal-time speech-to-text
    Apple Speech · patientReal-time speech-to-text
  3. Speaker-tagged transcriptAssembled and kept on the device.
  4. Foundation Model + clinical LoRA adapterNote summary on device, plus text-to-speech playback.no patient audio or transcript leaves the phone
  5. Node.js backend · Google App EngineAuth and sessions. It coordinates the workflow but never receives patient audio or transcripts.

Outcome

100%AI processing on device
6+ reposiOS, Android, shared Swift package, backend, built solo
PilotLimited production use with early customers

What I'd do differently

Build a scored evaluation set of consented, de-identified clinical phrases before the first fine-tune, so every adapter version is measured rather than judged by feel.

Swift · SwiftUI · Apple Foundation Models · LoRA · Apple Speech · Twilio Video · Android · Node.js · TypeScript · Google App Engine
04 / NEOM Smart City

One framework under ~20 apps, so no team rebuilt networking or auth.

Senior / Lead iOS Engineer · NEOM via Tech Mahindra · Jun 2023 – Jan 2024 · Saudi Arabia

NEOM was building many apps for residents, visitors and service providers at once. Without a shared foundation, every team would have built its own networking, auth and security, each slightly different and each patched separately.

Constraints

Options I weighed

The design

  1. Core framework · SPM, versionedNetworking · OAuth 2.0 · SSL pinning · Apollo iOS for type-safe GraphQL · shared UI
  2. Adopting apps
    Smart City appapp 2app 3…~20 apps
  3. Smart City app featuresIn-app assistant for utilities, transport and support (OpenAI client with LangChain-based prompts), Twilio audio and video for resident-to-agent support, and Apple Speech for voice.

Outcome

~20Internal apps adopted the core framework
v1The Smart City app's first App Store release
3Engineers mentored to launch

What I'd do differently

Ship a small sample app alongside the framework from week one. Getting 20 teams to adopt something is mostly a documentation problem.

Swift · SwiftUI · Swift Package Manager · Apollo iOS · GraphQL · OAuth 2.0 · SSL pinning · OpenAI · Twilio · Apple Speech
05 / Kroger

The release pipeline the whole iOS org shipped with.

Senior / Lead iOS & Release Engineer · Kroger via Tech Mahindra · Jan 2023 – Jun 2023

Kroger runs a family of grocery banners (Kroger, Ralphs, Fred Meyer, Harris Teeter), and each is a separate App Store app with its own signing and release. I was also delivering a new customer-facing feature, Gamification, to all of them.

Constraints

Options I weighed

The design

  1. Pull request → GitHub ActionsEvery change goes through the same path.
  2. Fastlane lane (banner, environment)
    krogerralphsfred meyerharris teeterdev | prod
  3. Build → sign → deliverPer-banner identity, then TestFlight or the App Store.
  4. Feature flags + experimentsFirebase Remote Config flags and A/B tests gate the Gamification rollout.
  5. MonitorCrashlytics and Grafana, watched during each rollout.

Outcome

4Banner apps on one pipeline
Org-wideThe pipeline was adopted by the entire iOS org
v1 → v2Gamification shipped to the full customer base

What I'd do differently

Add automated screenshot diffs per banner. Most multi-brand bugs are visual, like the wrong logo or colour, and only humans were catching them.

Swift · Objective-C · SwiftUI · Fastlane · GitHub Actions · Ruby · shell · Node.js · Firebase Remote Config · Crashlytics · Grafana