Case Studies

Challenge: Research across a large set of YouTube channels produces exactly the kind of data that text answers handle badly — rankings, comparisons, publishing frequencies, transcript segments. An assistant can fetch it all, but reading the results back as prose is slower than the manual work it replaced. The bottleneck was never retrieval. It was review.

Solution: A production MCP server built on the official TypeScript SDK, exposing 31 tools across channel analytics, video search, transcript extraction, and keyword indexing, with Supabase-backed persistence so results accumulate instead of being re-fetched.

The distinguishing part is the interface. Seven tools declare MCP Apps Views — HTML interfaces served as ui:// resources, which the tool advertises through _meta on its listing so the AI client renders the result inline rather than describing it. Analytics, search results, channel comparisons, and timestamped transcripts each get a purpose-built view instead of a paragraph.

Key Features:

• 31 tools spanning channel analytics, search, transcripts, and keyword indexing
• Seven interactive MCP Apps Views served as ui:// resources
• Shared core behind both an MCP server and a standalone CLI — scriptable without an AI client in the loop
• Supabase persistence, error reporting, and channel exclusion lists
• Dedicated host harness and automated test for the UI build, so views are verified rather than eyeballed

Results:

• Running in production on my own content research, not a demonstrator — 50+ commits and actively maintained
• Runs across multiple MCP clients — the same server is registered and enabled in both Claude and Goose, which is the portability claim of the protocol actually put to the test
• Comparative results are read, not parsed — rankings and channel comparisons arrive as a table rather than a paragraph, which is the whole point of the View layer
• The pattern generalises — the same View mechanism can render a proposed write as a reviewable diff, which is how you would let an assistant act on business data safely

Technologies: TypeScript, Model Context Protocol, MCP Apps (ext-apps), Supabase, YouTube Data API, Vite

Challenge: AI assistants lacked direct integration with Gamma, a popular presentation platform. Users had to manually copy content between tools, breaking their workflow and reducing productivity.

Solution: Took an early open source MCP server and extended it into something production-usable, connecting AI assistants (Claude Desktop, Cursor, etc.) directly to the Gamma API so AI can create, update, and manage presentations programmatically. 17 of the repository's 21 commits are mine.

Key Features:

• Create new presentations with AI-generated content
• Update existing slides programmatically
• Search and retrieve presentation data
• Full TypeScript implementation with comprehensive error handling
• Complete documentation and usage examples

Results:

• Used on live client work — the same Gamma integration produced a 15-slide executive deck and a detailed intelligence report for a consultancy's company and industry research engagement, generated straight from the analysis documents
• Open sourced on GitHub under my own fork, benefiting the wider MCP community
• Presentation generation became a tool call — no copying content between the assistant and the deck
• A reusable pattern for wrapping any commercial REST API as clean MCP tools with proper credential handling

Technologies: TypeScript, Model Context Protocol, Gamma API, npm packaging

View on GitHub →

Challenge: Developers struggling to choose between rapidly evolving AI tools, frameworks, and services. Information scattered across blog posts, documentation, and marketing materials made objective comparison difficult.

Solution: Created a curated comparison framework documenting AI tools, coding assistants, and development frameworks. Structured tables enable side-by-side feature comparison with objective criteria.

Key Features:

• Side-by-side comparisons of major AI coding assistants
• Framework evaluations (CrewAI, LangChain, AutoGen, etc.)
• Updated regularly as tools evolve
• Community-driven with contribution guidelines
• Markdown format for easy updates and version control

Results:

• 809 commits and still maintained — updated through 2026 as the tooling landscape shifts
• MIT licensed and open to contribution, with published contribution guidelines
• Helps developers make informed decisions based on objective criteria rather than vendor marketing

Technologies: Markdown, Documentation, GitHub

View on GitHub →