framlit
LiveEditable AI-generated Remotion motion graphics — generate a video, then tweak it as clickable layers.
- Next.js
- Remotion
- Claude
- AWS Lambda
AI engineer & builder
I build AI-native software,
and I ship it
relentlessly.
Roughly 80 products since December — motion-graphics engines, agent-native CLIs, data tools, and full SaaS apps. Many of them are live. By day I'm an AI engineer in IBM Client Engineering, turning enterprise problems into deployed AI systems; nights and weekends I turn ideas into shipped products.
Featured work
The ones that are live, polished, or just my favorites.
vibeframe
LiveTurn a written brief into a rendered MP4 using a coding agent.
- TypeScript
- HyperFrames
- MCP
- FFmpeg
querypad
LiveCursor for Data — a local-first AI workspace that understands your datasets.
- TypeScript
- DuckDB
- Next.js
- Claude
reframe
LiveDeclarative motion graphics AI can write, humans can tweak — and the human edits survive an AI regeneration.
- TypeScript
- GSAP
- FFmpeg
- Playwright
contextix
LiveCLI toolkit for agentic AI — ingest any source into a queryable, typed knowledge graph.
- TypeScript
- MCP
- SQLite
- Claude
mocksnap
LivePaste JSON or drop an OpenAPI spec — get a live REST + GraphQL API with docs in seconds.
- Next.js
- Hono
- GraphQL
- Claude
parsemux
LiveOne unified interface for five document parsers — auto-routed to the best engine per file.
- Python
- FastAPI
- MCP
- Docker
pitchframe
BetaA rough brief becomes a HeyGen-presented product video — real presenter plus motion graphics.
🥇 HeyGen Hackathon — 1st place
- Next.js
- HeyGen
- HyperFrames
- OpenAI
ponder
BetaA flagship build — add a one-line description and live link.
Coming soon
Selected builds
A slice of the rest, by theme. There are more where these came from.
Video & Motion
10- typeframe Spoken video in, stylized typing-caption MP4 out. Beta
- reframe-studio Motion-graphics Cursor: natural-language edit → semantic diff → accept or iterate. Beta
- stereoframe Declarative <sf-*> HTML markup → seekable 3D MP4 plus inspectable GLB. CLI
- framecode Source-first creative IDE for motion design, image, and video workflows. Prototype
- songscript Prompt-to-editable music markup with deterministic stem diffs — change one stem, rerender only that one. CLI
- traceframe Model architecture → motion graphics, straight from ONNX / PyTorch / Keras metadata. CLI
- timeframes Data-driven bar-chart races for long-run time series — a 1900–2026 GDP race. Prototype
- mermaid-to-hyperframes Mermaid flowcharts → seekable GSAP motion video. CLI
- strokeframe TeX in, chalkboard MathJax motion out. Prototype
- vibestudio Local-first OSS video studio for building AI video sequences. Prototype
Agent Tools & CLIs
10- snact Snap + act — AI browser control with extreme token efficiency, in Rust. Live
- cadgate Validate AI-generated CAD-as-code PRs — geometric diff, DFM rules, rendered previews, LLM judge. Live
- screenagent Control any macOS app with Claude — a Python SDK + CLI. CLI
- pagewise URL to QA in one CLI — scrape, chunk, embed, and query any webpage with hybrid RAG. CLI
- rawtowise Turn raw documents into a searchable knowledge wiki with [[backlinks]] and incremental compiles. CLI
- datacontext Make AI understand your data better — an AI-native database context layer over MCP. CLI
- agentic-cli-guide Seven principles for designing CLIs that AI agents can actually use well. Prototype
- morphframe AI-native adaptive frontend runtime — deterministic policy + replay, never raw generated HTML. Prototype
- skillforge Create, manage, and distribute Claude Code skill plugins. CLI
- agenttrust A trust & discovery layer for the agent economy — registry, benchmarks, certification. Prototype
Data & Analytics
4- vibe-lakehouse A $0 lakehouse: Cloudflare R2 + Apache Iceberg + DuckDB, queried in natural language by Claude Code. Prototype
- trendcontext AI-native trend intelligence from Naver, Google News, and YouTube. Prototype
- media-brief PR narrative-risk intelligence — daily briefs on narrative shifts with an evidence graph. Prototype
- contextix-platform Open-source world context for AI agents — structured causal understanding of world events. Prototype
Apps & SaaS
6- incomelens See where your salary ranks — compare pay and cost of living across US cities, plan for life changes. Live
- refit-aura AI fashion try-on — see outfits on you before you buy them. Beta
- prismdeck A local AI document → presentation desktop app. Prototype
- verideck Turn academic papers into editable presentation slides. Prototype
- pulseoverlay JSON in → live motion overlay out — a brandable OBS browser source from any endpoint. Prototype
- historacle A time-machine debate — seven historical thinkers argue modern issues. Prototype
Korea-focused
6- dart-fss-cli All 83 Korea FSS DART disclosure endpoints from one agent-native CLI. CLI
- ncli Agent-native CLI for Naver Open APIs — schema introspection, hardened input, NDJSON output. CLI
- company-pulse 이 회사 가도 될까? — AI corporate due-diligence from DART filings and the web. Prototype
- streampick 흩어진 OTT를 한눈에 — a unified guide across every Korean OTT. Prototype
- transitlens Korean public-transit CLI — bus and subway info with transfer pathfinding. CLI
- hanlytics Korean YouTube comment intelligence — sentiment, topics, and content strategy. Prototype
…and roughly 35 more experiments, prototypes, and half-finished ideas not listed here.
About
I'm an AI engineer on IBM's Client Engineering team in Korea, where I take enterprise problems from discovery and architecture through hands-on builds to production — RAG pipelines, multi-agent systems, and the evaluation harnesses that keep them honest.
Since December I've been building in the open at a different tempo — shipping small, sharp products almost weekly. A few themes keep recurring: determinism (renders you can reproduce frame-for-frame), edit survival (human tweaks that outlast an AI regeneration), and tools designed for agents first. The motion-graphics engines, the agent-native CLIs, and the data tools above all come out of that.
Before IBM I did graduate ML research at UT Austin — statistical learning theory and high-dimensional optimization — which is where the instinct for rigorous evaluation comes from. I work in Korean and English, and I care most about the unglamorous part of AI: making systems people actually trust and use.