Lex Hrabovskyi

Problem

Talking-head videos need instant, accurate word-level captions; cloud speech recognition adds latency, cost, and privacy trade-offs.

Constraints

Everything runs on device: real-time ASR under tight CPU and memory budgets, a 60 FPS editing UI, and exports that must match the in-app preview frame-for-frame.

My role

Sole engineer and product owner — ASR engine, module architecture, the capture → align → edit → export pipeline, and profiling.

Result

≈180 ms word-follow latency at ≈13% CPU / 40 MB with no cloud calls; playback main-thread CPU cut 41.4% → 35.2% and peak memory 115 → 99 MB at a held 60 FPS.

Business impact — Zero inference cost and full privacy — the product works offline, end to end, on the phone.

Highlights

• Built LexSpeech, an on-device ASR engine reaching ≈180 ms word-follow latency at ≈13% CPU / 40 MB — no cloud calls.
• Architected the app as 14 layered SPM modules (Foundation → Services → Features → AppHost) with compiler-enforced import boundaries and Factory DI, so every feature builds and tests independently.
• Cut combine-playback main-thread CPU from 41.4% to 35.2% and peak memory from 115 MB to 99 MB while holding 60 FPS — profiling with Instruments and windowing the word strip to on-screen cells.
• Developed the end-to-end capture → align → edit → export pipeline (SwiftUI + SwiftData, iOS 26) with animated per-word captions — the in-app preview matches the exported video frame-for-frame, and stitched takes join within one frame.

In active development — not yet on the App Store.

Problem

Trip detection quality directly affected user trust in the core product.

Constraints

Legacy Objective-C code, sensor variability, background execution, regression risk, and release safety.

My role

Owned the trip-detection engine, evaluation loop, rollout safety, and analytics.

Result

Reduced split trips by ~50% and achieved sub-20-second trip starts using signal-driven warm-up.

Business impact — Improved reliability of the product’s core experience without destabilizing production releases.

Highlights

• Cut split trips by ≈50% and achieved sub-20-second trip starts by introducing signal-driven warm-up.
• Prevented regressions before release with replay-based integration tests around the legacy Objective-C stack; staged rollouts behind feature flags.
• Set up analytics — event schema, trip-funnel dashboards via Bitdrift — and partnered with the data team to evaluate engine changes.
• Planned a safe rewrite: a six-module design and Swift 6 upgrade path that lets the engine evolve without breaking releases.

Problem

A live retail app stuck at 3.4 stars with stability issues — ratings and reviews were hurting a brand with 150+ physical stores.

Constraints

Large legacy codebase, a monthly release train that could not slip, and modernization that had to land without breaking production.

My role

Led iOS delivery and the modernization strategy; coordinated QA and design, mentored developers, and ran a parallel internal product initiative.

Result

Rating up from 3.4 to 4.6 at ≈99.8% crash-free; codebase modernized with SPM feature modules, DI, and async/await while sustaining monthly releases.

Business impact — The App Store rating went from liability to asset for the retail brand, and the modular codebase made delivery measurably faster.

Highlights

• Raised the app from 3.4 to 4.6 stars at ≈99.8% crash-free through targeted bug fixes, UX improvements, and an in-app review flow.
• Cut legacy complexity and sped delivery by introducing SPM-based feature modules, dependency injection, an async/await migration, and selective SwiftUI adoption.
• Hardened quality and CI with XCTest, XCUITest, and snapshot testing, a mock API server, and preset environments for GitLab CI UI runs — sustaining a monthly release cadence.
• Led a parallel internal product initiative: owned scope and roadmap, coordinated QA/design, mentored developers, and landed a critical mid-cycle release.

Highlights

• Delivered an AR session-recording pipeline (ARKit, Metal, AVFoundation) capturing video, depth, and 3D objects into a single package, with on-device post-processing for 60 FPS replays.
• Streamed strokes over WebSockets to a real-time 3D iPad mannequin preview, reusing the rendering core to avoid duplicate logic.
• Shipped a SwiftUI revamp of ≈12 screens in 2 months, removing unused 3D UI from an ARKit scene to unblock a customer contract.

Distributed via TestFlight to enterprise customers — no public App Store listing.

Highlights

• Cut memory ≈8× (406 MB → 50 MB) via image caching, downsampling, and cleanup.
• Built CI/CD from scratch on Bitrise with signed TestFlight builds and staging/prod presets.
• Shipped a crypto wallet from scratch to the App Store and built Spotify-SDK social-listening features.