LLM Odyssey 2025

Local LLM TUI + Multi-modal Models 👀

• Built OllamaTea - BubbleTea components for Ollama
• Terminal UI meets local LLM inference
• Combined with dbn-go DataBento market data
• Created NTCharts multi-modal inferencing demo

Industry Milestones

MCP Exploration 🤖

• Built MCP servers with mcp-go:
• dbn-go-mcp for DataBento market data
• ████plug-mcp for device access
• buttplug-mcp for device access
• dink-mcp: connecting Claude with structured datasets via DuckDB SQL
• dank-mcp: connecting Claude with cannabis datasets via DuckDB SQL

Self-improvement of MCP Server via Agentic Loops 🔄

World knowledge + SQL generation = deterministic answers to probabilistic questions

• Built screentime-mcp — querying macOS ScreenTime database (SQLite)
• Key pattern: Tool Description Markdown 🔨 and Preflight SQL ⚙️ (Views and Funcs)
• 🥹 Human: noticing patterns, updating files — added hot reloading
• 🤖 Agentic loop: LLM self-improved MCP files through my personal self-improvement prompting

Take a Shot at Vibe Coding 👎

• Vibe coding was getting popular
• Already using Copilot auto-complete heavily
• Constantly asking LLMs coding questions
• Tried making AgentDink website — starting with age verification
• Tried making AgentDank website — starting with age verification
• Didn't concentrate on the code at all...

AntiGravity + Frontier Models 😵‍💫

• Gemini 3 Pro + Claude Opus 4.5
• One-shot: AgentDink verification page + chat interface
• One-shot: AgentDank verification page + chat interface
• Then free to vibe new features...
• One-shot: DBA environment with embedded DuckDB WASM
• Then vibed 3D WebGL charts of query results
• Built between meetings and commutes

Vibe Coding for Kids 😵‍💫

• Substitute teaching - kids on Google Classroom
• Live coded on the smartboard:
• Quadratic Formula Explorer (repo)
• Proving Parallelograms (repo)
• Sprinkled inspiration about machine intelligence

Multimodal Debugging

See image → Identify bug → Create fix → Write test → Run in Chrome → Screenshot → Report

Vibe Boat Engineering 🚤

• Multi-modal mechanical design assistance for pontoon boat upgrade
• Described issues + shared CAD drawings
• AI helped with:

Structure configuration • Screw diameters • Materials selection • Stress point analysis

Image → Swift Code → 3D Model → Image...

• Discovered Cadova, a Swift library for parametric, 3D CSG modeling
• Realization: Swift as ground truth for 3D models
• Like how Claude generates DOCX or SVG... this is Swift code → 3MF model
• Documented in WeaselToonCadova repo

LLM-Assisted Renaissance 🚀

• Now I am jacked in with multiple projects simultaneously
• With multiple LLMs, multiple workflows, multiple tools

• NTCharts multimodal + terminal testing
• HyperCube Viewer on Vision Pro via Mesh Shading
• AgentDink dink-extract
• AgentDank dank-extract
• Swift app for WeaselToonCadova loop
• OSS devops: Travis → GitHub Actions
• BuffaloBuffalo
• This Presentation... (PROMPTS.md)
• Business + financial planning

Artistic License 🎨

• Wanted to make something Fallthrough inspired
• Start with photo of hosts → LLM iterates Swift 🔄 3MF → optional human prompt → 3MF
• Fine-tune 3MF on additional reference, with face detection 🔄
• Generated 3D model → Printed It!