Projects

A journey through the products, platforms, and intelligent systems I've designed, engineered, and brought to production.

AI SaaS · Developer Tool

Astra

Role: Founder & Full-Stack Engineer

May 2026 – Present

Astra reads your actual codebase through AST-level analysis across 20+ programming languages, auto-detects frameworks and architecture patterns, then generates AI-crafted technical narratives tailored for engineering audiences — deployed live in under 60 seconds. No config files. No manual descriptions. Just your GitHub, as a portfolio.

Key Contributions & Technical Impact

Built an AST-level code analyser supporting 20+ languages that infers architecture patterns, framework usage, and design decisions directly from source — no README required.
Designed a GitHub webhook pipeline achieving under 1-minute median sync lag so portfolios auto-update on every push without user intervention.
Engineered a Gemini-powered narrative layer that produces technical descriptions tuned for engineering audiences, replacing the manual 'describe your project' step entirely.
Implemented one-click GitHub OAuth with read-only scopes and zero persistent token storage — auth is stateless by design to minimise the attack surface.
Shipped three distinct portfolio renderers (Minimal, Terminal, Void) sharing a single content graph, with shareable deep-links per project and per-repository editorial override controls.

Next.jsTypeScriptRustGoogle GeminiGitHub OAuthGitHub WebhooksAST Analysis

2026 – PresentAI SaaS

AI SaaS · WeCommit

Fundscout

Role: Full-Stack AI Engineer

April 2026 – Present

Production-grade AI platform (WeCommit SBIR) that automates federal SBIR/STTR grant discovery and proposal generation for small businesses. Features persistent deep agent conversation memory, real-time multi-user document editing via Yjs CRDTs + Hocuspocus WebSockets, and custom Gemini model routing.

Key Contributions & Technical Impact

Raised production uptime to 99.2% by diagnosing async PostgreSQL pool exhaustion under concurrent agent load — exponential backoff retry states fixed what connection pool tuning alone couldn't.
Cut grant proposal generation to under 3s for 50KB documents — the bottleneck was mapping LangGraph's event stream to Vercel AI SDK DataStreamContext; direct SSE caused multi-user session collisions that this solved.
Achieved under 100ms real-time document sync via Yjs CRDT + Hocuspocus WebSockets; the hard part was reconciling CRDT merge operations with mid-session AI content injections without losing user edits.
Architected a 5-layer middleware deep agent stack (Todo, Filesystem, SubAgent, Skills, Memory) with PostgreSQL checkpointing for long-running agent resumption across disconnected sessions.
Built custom GeminiRoutingModel supporting runtime 2.5 Pro vs Flash selection and dynamic thinking budgets — model routing alone cut average response cost by routing simple queries to Flash.
Developed custom Tiptap 3 editor extensions for inline artifact embeds and real-time AI content injection without corrupting document schema.

React 19TypeScriptTiptap 3FastAPILangGraphGemini AIPostgreSQLYjsHocuspocusAlembicPlaywright

AI SaaS · WeCommit

StageX

Role: Technical Lead & Core Architect

March 2026 – Present

AI-powered startup execution SaaS platform serving 200+ founders during an 8-week accelerator program. Integrates dynamic AI planning with LangGraph agent co-founders, a visual React Flow canvas, real-time streaming, and stateless mentor coordination loops. Deployed live on https://stagex.kr.

Key Contributions & Technical Impact

30+ founders onboarded across an 8-week accelerator — cut roadmap + full report generation from 180s+ to under 30s by shifting from sequential LLM calls to Vercel AI SDK streaming with parallel agent steps.
Engineered dynamic multi-agent SaaS orchestration using LangGraph and Vercel AI SDK; progressive streamObject streaming eliminated blank-screen waits on long-running plans.
Designed a 4-node deep research agent (SearchQuery → WebSearch → Synthesis → Report) with a pre-deduction credit ledger — users aren't charged for failed or cancelled runs.
Built an interactive tree layout engine with React Flow for drag-drop roadmap sync; the challenge was reconciling local drag state with the PostgreSQL-persisted graph in real time.
Mentor auth first used JWTs in localStorage — switched to httpOnly cookies after identifying XSS exposure, then added dynamic cohort filtering and dashboard feedback UI.
Integrated Paddle subscription billing with idempotent webhook event handlers to prevent duplicate charge processing under retry storms.

Next.js 14React 18TypeScriptLangGraphReact FlowPrismaPostgreSQLPaddleNextAuthVercel

2026 – PresentAI SaaS

2025 – 2026AI SaaS

AI SaaS · WeCommit

GenDoc (젠닥)

Role: Lead Full-Stack UI & Integrations Engineer

March 2025 – January 2026

AI-powered government grant automation and enterprise document collaboration platform designed for startups and SMBs. Streamlines grant discovery, automates proposal drafting matching complex rubrics, and hosts a secure, collaborative rich-text office editor with full revision rollbacks.

Key Contributions & Technical Impact

Deployed to Fly.io Tokyo to achieve under 50ms API latency for Korean users — multi-stage Docker builds kept image size lean without sacrificing runtime dependencies.
Reverse-engineered Collabora Online's undocumented WOPI save protocol to build a fully compliant Express.js WOPI Host; concurrent multi-user saves required custom timestamp-based collision locking headers that Collabora's docs don't mention.
Built S3 snapshot timeline with automatic version history and zero-data-loss restoration — the challenge was atomically coordinating S3 object versioning with document state in PostgreSQL to prevent partial restore scenarios.
Engineered a responsive React workspace with Radix UI and Tailwind CSS v4; Framer Motion transitions were tuned to the collaboration UX rhythm so state changes feel instant, not janky.
Bridged Naver/Kakao OAuth non-standard token formats to the Better Auth adapter interface — Korean providers don't follow OAuth 2.0 spec uniformly, which required custom normalisation middleware.
Orchestrated multi-stage Docker builds across Fly.io production and Vercel staging previews, with GitHub Actions gating staging deploys before Tokyo promotion.

ReactExpress.jsTailwind CSS v4Better AuthAWS S3Drizzle ORMPostgreSQLDockerFly.ioVercel

AI · Image Generation & Community Platform

CROSS-E

Role: Full Stack Engineer

August 2024 – September 2024

A creative showcase utilizing advanced generative AI models to convert text prompts into high-fidelity custom images. Integrates a vibrant community feed, real-time social sharing, search index filters, and instant media downloads.

Key Contributions & Technical Impact

Direct DALL-E API calls caused request timeouts under concurrent load — rebuilt around an async task queue with MongoDB job tracking, cutting generation failure rate to near zero.
Architected media storage on AWS S3 with CloudFront CDN for global image distribution; cache invalidation strategy was designed to minimize egress costs on re-generation.
Developed a lightweight community feed with Express.js and MongoDB — schema design prioritised read performance for paginated gallery queries over write flexibility.

React.jsNode.jsExpress.jsMongoDBDALL-EAWS S3AWS CloudFront

2024AI