Jay Heo

Building Ecco's cross-platform streaming app across web, mobile, and an airline in-flight entertainment build.

• Moved AI Search, the product's most differentiated feature, into a permanent navigation tab on mobile web and the mobile app, after event tracking I built showed how few users were finding it. First-month reach went from 20% to 25% of active users, repeat engagement more than doubled (24% to 54% returning 3 or more times), and daily core-area engagement climbed from 9.5% to 34% — holding after marketing spend dropped to zero.
• Ported the web app to airline in-flight entertainment hardware (an old Qt browser with barely any CPU or memory to spare) and built an MQTT-based remote so a separate device could drive it. Demoed at the Aircraft Interiors Expo, leading to follow-up meetings with 18 airlines including Lufthansa, Etihad, and Saudia.
• Led the React Native upgrade from 0.76 (past end-of-support) to 0.84 alongside React 19 and Expo 55, driven by Android 16 KB page-size requirements for Google Play and a native dependency stack (expo-video, FlashList v2, Reanimated v4) moving New-Architecture-only — using ~490 existing tests to catch regressions across iOS and Android.
• Rebuilt voice search around a proper speech-recognition state machine — stable vs. interim transcripts, intentional stops vs. dropped connections, rate-limited auto-restart — driving a 30% increase in searches after launch.
• Diagnosed and fixed a rare race condition in the video feed that only reproduced under rapid scrolling, where the index from user events and the one TanStack Virtualizer managed internally could drift apart and play the wrong video. Fixed by centralizing feed state in Zustand.
• Built the Go backend and Redis-based session logic for the analytics event pipeline (events that require a user_id vs. ones fired pre-login), and shipped the app's first monetization path: a Stripe subscription flow covering checkout, subscription status, and webhook-driven entitlements.

• Migrated the backend from REST to GraphQL/Apollo so the frontend fetches exactly what each resume section needs instead of over-fetching across multiple endpoints.
• Built the AI resume generation flow — calculating available line space before generation, with keyword tracking so bullets don't repeat across a resume.
• Moved resume file storage off the database and onto S3 with pre-signed URLs for previews and downloads.

• Cut AI response latency from 5–10 seconds to near real-time: the original pipeline recorded full audio, transcribed it with Whisper, then called the OpenAI API. Rebuilt it to stream audio over WebSockets in chunks and stream responses back token by token, after benchmarking faster inference like Groq.
• Built the real-time mock interview flow with Next.js dynamic routes talking to the AI backend over WebSockets, and split coding questions from general prompts before they hit the AI so responses came back more relevant.
• Coordinated project timelines and sprint planning with a Kanban workflow to hit each milestone ahead of Y Combinator's application deadline.

• Built a portfolio view pulling a user's on-chain holdings and calculating average cost basis, current value, and unrealized gains/losses against live prices — caching CoinGecko responses and polling on refresh to stay under a low API rate limit.
• Added QR-code actions tied to the user's wallet address, with options depending on what the scanned code was.