I went from running a restaurant to driving growth at SaaS companies to building cloud and AI systems. Not aspiring — already shipping. Everything I build is production-deployed, architecture-first, and designed to solve real problems for real users.
I spent years manually scheduling restaurant staff every Sunday night. Spreadsheet open, stack of availability texts, too much coffee. I built Schedio because I lived the problem.
Live: schedio.cloud — Free plan, no credit card required
Schedio connects to Google Drive, reads employee availability, and generates a full weekly schedule using AI. Managers adjust shifts via drag-and-drop or by chatting with an AI assistant in natural language.
Tech Stack
| Layer | Technology |
|---|---|
| Frontend | Next.js 14, Tailwind CSS, Vercel (edge CDN) |
| Backend | FastAPI (Python), SQLAlchemy |
| Database | Supabase (PostgreSQL + pgvector) |
| AI | OpenAI gpt-4o-mini, RAG pipeline, streaming chat |
| Infrastructure | AWS ECS Fargate, ALB, WAF, Route 53, Secrets Manager, ECR |
| IaC | Terraform — full stack deployable with terraform apply |
| CI/CD | GitHub Actions — build, push to ECR, deploy to ECS on every push to main |
| Payments | Stripe Checkout + Customer Portal |
Highlights
- RAG architecture with intent detection — queries automatically routed to the right context (availability vs. schedule history)
- Streaming AI chat — responses appear in real time with confirm/cancel flow before any schedule change is applied
- Multi-tenant data scoping at the database level — all data isolated by
user_id + location_id - LLM abstraction layer — swap between OpenAI and AWS Bedrock via a single env variable
- Google Drive OAuth integration for availability ingestion
- Freemium model with three plan tiers enforced at the prompt, shift, and database level
- 7+ production deployments
Small businesses lose wrongful termination cases because they have no paper trail. ShiftScore makes daily performance documentation feel as natural as swiping TikTok.
Live Demo: demo credentials in repo
A mobile-first performance tracking platform for shift-based businesses — built so supervisors can rate their team in under two minutes per shift, and managers can build a legally defensible, data-backed performance record. Behind the scenes: radar charts, trend lines, real-time alerts, and AI-generated weekly report cards delivered by email.
graph TB
subgraph Client["Mobile Browser — Next.js 16"]
Swipe["Framer Motion\nSwipe Gesture Engine"]
Charts["Recharts\nRadar · Heatmap · Trend"]
RT["Supabase Realtime\nWebSocket — alerts only"]
end
subgraph API["Next.js API Routes"]
Guard["JWT Auth + Role Guard\nproxy.ts"]
Routes["/api/sessions\n/api/ratings\n/api/alerts"]
end
subgraph Data["Supabase"]
PG["PostgreSQL\nRLS on every table"]
Realtime["Realtime\nalerts table"]
EdgeFn["Edge Functions (Deno)\npg_cron scheduled jobs"]
end
subgraph AWS["AWS"]
SES["SES\nTransactional email"]
Bedrock["Bedrock — Claude 3 Haiku\nAI performance narratives"]
CW["CloudWatch\nStructured logs"]
end
Client --> Guard
Guard --> Routes
Routes --> PG
Realtime --> RT
EdgeFn -->|"Mon 8AM — pg_cron"| SES
EdgeFn --> Bedrock
Tech Stack
| Layer | Technology |
|---|---|
| Framework | Next.js 16 (App Router), TypeScript (strict) |
| Styling | Tailwind CSS |
| Animation | Framer Motion — swipe gesture engine with spring physics |
| Charts | Recharts — radar, line, heatmap |
| Database | Supabase (PostgreSQL + RLS) |
| Auth | Supabase Auth — role resolved from DB on every request, not JWT claims |
| Real-time | Supabase Realtime — Postgres change events scoped to alerts table |
AWS SES with full email_log audit trail |
|
| AI | AWS Bedrock (Claude 3 Haiku) — weekly AI narrative report cards |
| Scheduler | Supabase Edge Functions (Deno) + pg_cron |
| Deployment | Vercel |
Highlights
- Three-role system (owner / supervisor / employee) enforced at the proxy layer — role is resolved from the database on every request, not from JWT claims, so access changes take effect immediately without re-login
- Swipe gesture engine runs on
useMotionValue— gesture state never enters React, only a committed swipe triggers a state update - Real-time alert feed: consecutive-poor-rating detection triggers a Postgres change event → Supabase Realtime → admin browser with zero polling
- Weekly AI report cards: Deno Edge Function runs Monday 8AM via pg_cron, calls AWS Bedrock per employee, sends via SES, logs every message ID for audit
- RLS enforced on all tables — a compromised anon key cannot read cross-role or cross-location data
- Upsert-based ratings API — idempotent by design, re-submitting a changed swipe updates the existing row rather than duplicating
111 sources. Scraped daily. AI-enriched. Served globally from CloudFront. Zero manual steps.
Live: novakidlife.com
A production, monetized web platform that aggregates family-friendly events, deals, and Pokémon TCG activities across Northern Virginia. The entire backend is an automated data pipeline — scraping, image generation, blog content, and site deployment all run on schedule without any manual intervention.
flowchart TD
subgraph Pipeline["Automated Data Pipeline"]
EB["EventBridge\nCron Scheduler"]
Scraper["events-scraper Lambda\n111 sources · 3-tier + Pokémon"]
SQS["SQS + DLQ"]
ImgGen["image-gen Lambda\nGoogle Places → Unsplash → Pexels → Imagen 3"]
Content["content-generator Lambda\n5 post types · 2×/week"]
end
subgraph Infra["AWS Infrastructure (Terraform)"]
CF["CloudFront CDN\nACM wildcard SSL"]
S3Web["S3 Static Web\nNext.js SSG output"]
S3Media["S3 Media\nWebP variants + LQIP"]
APIGW["API Gateway\napi.novakidlife.com"]
Lambda["API Lambda\n15 routes · Python 3.12"]
SSM["SSM Parameter Store\n18+ SecureString secrets"]
end
subgraph Data["Supabase"]
DB["PostgreSQL + pgvector\nSemantic search index"]
end
GH["GitHub Actions\n5 CI/CD workflows"]
EB -->|"Daily 6AM EST"| Scraper
Scraper --> SQS
SQS --> ImgGen
ImgGen --> S3Media
ImgGen --> DB
EB -->|"Thu 8PM · Mon 6AM"| Content
Content --> DB
Content -->|"trigger deploy via API"| GH
GH -->|"npm build + S3 sync + CF invalidation"| S3Web
CF --> S3Web
CF --> S3Media
CF --> APIGW
APIGW --> Lambda
Lambda --> DB
Tech Stack
| Layer | Technology |
|---|---|
| Frontend | Next.js 15 (static export), TypeScript, Tailwind CSS 3.4 |
| Hosting | AWS S3 + CloudFront (CDN + ACM wildcard cert) |
| API | AWS API Gateway + Lambda (Python 3.12, 15 routes) |
| Database | Supabase (PostgreSQL + pgvector) |
| Queue | AWS SQS + DLQ |
| Scheduler | AWS EventBridge cron rules |
| AI — Images | Google Imagen 3 (primary) · DALL-E 3 (fallback) |
| AI — Content | OpenAI gpt-4o-mini — blog generation, alt text, AI-assisted scraping |
| AI — Search | OpenAI text-embedding-3-small → pgvector cosine similarity |
| IaC | Terraform (S3 backend, DynamoDB state lock) |
| CI/CD | GitHub Actions (5 workflows) |
| Secrets | AWS SSM Parameter Store (18+ SecureString params) |
Highlights
- 3-tier scraper architecture: structured APIs (Tier 1), AI-extracted config-driven sources (Tier 2 — add a new source by editing one JSON file, no code changes), deal monitors (Tier 3)
- Image pipeline: sourced from Google Places → Unsplash → Pexels, with Imagen 3 as AI fallback, then Pillow warm-graded and exported as WebP with LQIP blur-up placeholders for zero layout shift
- Content-generator Lambda triggers a GitHub Actions workflow via the GitHub API after publishing new blog posts — the static site rebuilds and deploys automatically
- Semantic search backed by pgvector — no separate vector database, cosine similarity lives in the same Postgres instance
- All 18+ secrets stored in AWS SSM Parameter Store; loaded at Lambda cold start — nothing hardcoded anywhere
- Lighthouse CI enforces 90+ scores on every pull request
A first-generation Korean immigrant at a doctor's appointment shouldn't have to guess what the doctor said. Sori records it, translates it, and surfaces the most important information first.
Status: Phase 1 in active development — iOS
A React Native/Expo mobile app that helps non-English speaking Korean immigrants understand high-stakes English conversations — doctor visits, school meetings, legal consultations — by providing speaker-labeled transcription, Korean translation, and structured AI summaries.
graph TB
subgraph Mobile["iOS App — React Native / Expo"]
UI["Expo Router Screens"]
Audio["expo-av\nAudio recording + PCM streaming"]
ApiSvc["apiService\nAll external traffic proxied through backend"]
SubSvc["RevenueCat\nSubscription state"]
end
subgraph Backend["Backend — Node.js / Express on Railway"]
Auth["JWT Auth Middleware"]
Cap["Usage Cap Guard\npre-recording quota check"]
Queue["Job Queue\nMax 3 concurrent STT jobs per user"]
WS["WebSocket Server\n/api/stream"]
end
subgraph AI["AI Services"]
FriendliAI["FriendliAI\nwhisper-large-v3\nBatch STT"]
Deepgram["Deepgram Nova-3\nStreaming STT + Speaker Diarization"]
Claude["Claude API\nclaude-haiku-4-5\nTranslation + Summary"]
Pyannote["pyannote.audio\nSpeaker Diarization (Mode 1)"]
end
subgraph Data["Supabase"]
PG["PostgreSQL\nText only — audio never persisted"]
TempBucket["Storage: audio-temp\nDeleted immediately after STT"]
end
Mobile --> Backend
Auth --> Cap
Cap --> Queue
Queue --> FriendliAI
Queue --> Pyannote
FriendliAI --> Claude
Pyannote --> Claude
Claude --> PG
Queue --> TempBucket
TempBucket -.->|"auto-deleted post-STT"| TempBucket
WS --> Deepgram
Deepgram -->|"per sentence"| Claude
Two modes:
| Mode 1 — Record & Translate | Mode 2 — Live Translate | |
|---|---|---|
| Flow | Record → upload → batch process | Real-time WebSocket streaming |
| STT | FriendliAI (whisper-large-v3) | Deepgram Nova-3 |
| Latency | 2–4 min for a 30-min recording | <200ms English / ~1–2s Korean |
| Tier | Free + paid | Paid only |
Tech Stack
| Layer | Technology |
|---|---|
| Mobile | React Native + Expo SDK, TypeScript |
| Backend | Node.js + Express (Railway) |
| Database | Supabase (PostgreSQL + RLS) |
| Auth | Supabase Auth |
| STT (Batch) | FriendliAI — whisper-large-v3 |
| STT (Stream) | Deepgram Nova-3 (WebSocket) |
| Diarization | pyannote.audio |
| AI | Claude API — claude-haiku-4-5 |
| Subscriptions | RevenueCat |
| Analytics | PostHog (anonymous events only) |
| Crash Reporting | Sentry |
Highlights
- Audio is never stored permanently — written to a private Supabase temp bucket, deleted the moment STT completes, with a 24-hour auto-expiry safety net
- Claude prompt engineered to handle Konglish code-switching — Korean speakers naturally mix English medical/legal terms, and the translation preserves that pattern naturally (e.g.
혈압(blood pressure)이 높습니다) - Mode 2 reconnection logic: buffers the last 30 seconds of audio locally, retries the WebSocket every 3 seconds up to 5 times, then auto-falls-back to Mode 1 batch if all retries fail
- Usage cap enforced server-side before every recording starts — queries Supabase for
minutes_usedvs. tier limit, returns 402 if exceeded, never relying on client-side gating - All AI API keys live on the Railway backend — the mobile client holds only the Supabase anon key
- LinkedIn: linkedin.com/in/tae-joonkim
- Email: kim.taejoon@gmail.com