Keugene Lee

An AI agent that runs your job search end to end: it finds openings across 20,000+ company career pages, scores how well you fit, answers every application question, and submits from your own browser.

• Crawls Greenhouse, Lever, Ashby, Workday, SmartRecruiters, Workable and Recruitee through their public job APIs, 390,000+ open US and remote roles across 20,000+ boards, closing roles by id diff rather than timestamps and skipping the write entirely when a board's listing fingerprint hasn't changed
• Two-stage matching: a SQL prefilter and a batched LLM rank on title and location, then a deep score against the full description with reasons, gaps, and hard caps when you miss a real requirement. Your resume goes out exactly as you uploaded it
• A Chrome extension sends approved applications hands-off from pinned background tabs in a real browser, since Greenhouse, Lever and Ashby bot-score every submit, with exclusive server-side claims so nothing sends twice. A Workday driver keyed on data-automation-id creates per-employer accounts with AES-256-GCM-encrypted credentials, verifies them through Gmail and walks the multi-step wizard.
• Answers you can trust: deterministic rules decide sponsorship, work authorization, self-ID and conflict-of-interest questions before any model sees them, and a second model proofreads every application against your profile before it leaves, flipping anything that contradicts it. Every question and answer is logged and searchable
• When a send fails it is triaged, not blindly retried: the fields the form rejected are re-answered and resent, and repeated failures with one signature pause that job site until a fixed extension version checks in. Employer receipts are matched from Gmail every five minutes, so you can see which applications a company has confirmed
• A fast lane re-crawls the 1,800 boards that actually produce roles worth applying to every five minutes and applies on arrival, which cut the median age of a job at application time from 13 days to under an hour. Autopilot still owns the limits: one application per employer a month, plan and daily caps, under a cent of model cost per application

A personal AI agent — Claude Code for your life. Your first chat is an interview that builds a long-term memory of you; from then on the agent goes and does things: researches the web, finds places on Google Maps, checks weather and events, drafts messages, manages goals and tasks, and runs recurring routines you approve first

• Looks and works like Claude Code in a terminal: a monospace transcript of ● Tool(args) and └ result lines, a thinking spinner with live “Thought for Ns” timing, slash commands, light and dark themes, and effort modes from Fast to Standard, Deep and Marathon. No landing page — visitors land straight in the chat, and signing in runs the message they already typed
• A durable run engine so a task can run for up to 45 minutes on 300-second serverless functions: every run is a Postgres row, advanced by chained, lease-based worker invocations with crash recovery and streamed to the client over SSE. Parallel sub-agents run as child runs that wake the parent when they finish
• Inference routed across open-weight models — GLM-5.3-Flash via OpenRouter, then DeepSeek, then Groq — with rate-limit-aware retries and per-call cost tracking, so one provider hitting its limit doesn't stop a run
• iOS and Android apps through a Capacitor shell with native Sign in with Apple, Google sign-in, APNs push, haptics and Whisper voice input — built and signed in GitHub Actions, with TestFlight and Play internal testing builds uploaded
• Free and Pro plans ($20/month or $100/year) billed through Stripe on the web, StoreKit 2 on iOS and Play Billing on Android, with every purchase verified server-side instead of through RevenueCat — plus App Store privacy compliance, including explicit AI data-sharing consent enforced on the server

Always-on-top desktop assistant for Windows that can see your screen and use your computer — one hotkey reads whatever is in front of you and either answers it, walks you through it, or does it for you

• No mode switch: one streamed tool-calling request to a vision model (Gemini Flash-Lite via OpenRouter, DeepSeek as fallback, about a tenth of a cent per ask) decides whether the reply is an answer, a step-by-step walkthrough, or launching an app, and all three land in the same thread — the user never picks a tool
• Drives real input through a PowerShell host process — click, double-click, drag, scroll, typing, key combos, and restoring a minimized window — with a guiding cursor drawn on a full-screen click-through overlay so you can see what it is about to press. Every step declares what should be true on screen afterwards, gets re-checked against a fresh screenshot before the next one starts, and retries up to three times rather than claiming it worked
• Push-to-talk transcription runs entirely on-device — Whisper through transformers.js, imported from a CDN at runtime so it needs no extra API key and adds nothing to the installer — and answers are spoken back through the Web Speech API
• Desktop shell built on Electron with global hotkeys, a bar that resizes to hug its content and can be hidden from screen recording, and auto-updates via electron-updater. The UI is served from Vercel rather than bundled, so shipping a change never means shipping a new binary. Sessions end with a generated summary, key points, action items and a follow-up draft, stored in Neon behind JWT auth

AI music discovery app — enter a song you love and an LLM builds a playlist of tracks you'll actually want to hear, each with an explanation of why

• The model weighs genre, vibe, and sonic qualities using Last.fm, TasteDive, and ListenBrainz to match energy and mood, not just genre
• Every recommendation is cross-verified on YouTube and Last.fm so every song is real and playable
• Spotify-style playlist player with a now-playing bar, autoplay, and like/dislike feedback that shapes future recommendations

The Facebook for Stony Brook University — a social network where verified @stonybrook.edu students find classmates, connect by dorm, and stay in the loop, with 50+ registered users

• Full profile system with major, dorm, courses, Greek life, clubs, privacy controls, and profile-view tracking
• Wall posts with photo/video uploads, likes, threaded comments, friend requests, pokes, groups, and direct messaging
• Defense-in-depth security enforced in Postgres row-level security: email-gated signup, per-IP rate limiting, bot-pattern detection, and owner-protection policies — every privacy rule lives in the database, not the app

AI-powered day planner that turns a city and budget into a complete itinerary, pulling live data from Google Places, weather, transit, events, and more

• Backend calls every relevant tool in parallel, then Gemini Flash-Lite (via OpenRouter) synthesizes the results into real venues with prices, clickable Google Maps links, and an interactive map
• Watch the plan build live via Server-Sent Events, with a nightlife mode for evening plans and weather-based outfit suggestions
• Supabase auth with cloud-synced plan history