This is my AI-native portfolio. It's a SvelteKit app that knows everything about me and talks to you about it. Drop a question into the chat and it searches my blog posts, resume, and project docs before answering. It can also call real tools: browse my GitHub repos, check pull requests, pull photography from my Macula site.

Everything runs locally: embeddings, vector search, LLM calls. No external AI services beyond whatever model endpoint you point it at.

The core idea is pretty simple: a chat interface on top of everything I've built and written. When you ask something, the app chunks up my content into vectors, matches what you're asking against the index, then feeds the relevant bits to an LLM along with the tools it can call. You get answers backed by real content, not hallucination.

The LLM cache is worth calling out specifically. It's a USearch ANN index that checks if someone already asked something similar before hitting the API. Same question gets an instant cached response. Close question gets a partial match. Configurable threshold so you can tune how aggressive the dedup is.

Beyond the chat, there's:

• Blog and resume: markdown files with frontmatter, full-text and vector searchable
• OG images: rendered server-side per post with Satori and resvg-js, no external service
• Dark and light mode: follows your system preference, handled by sv5ui theming
• Rate limiting: SQLite-backed, 10 requests per minute per IP
• GeoIP: country detection via geoip-country so I know roughly where visitors are
• Structured logging: LogTape with file rotation plus a ZinaLog dashboard for browsing
• Webhooks: events pushed to external URLs when interesting things happen
• Message reactions: upvote, downvote, or heart AI responses
• Contact and intent detection: the AI can figure out if you're trying to hire me or just browsing
• Docker: multi-stage build, production ready

pnpm install
cp example.env .env
# Edit .env: set PROVIDER_API_KEY, LLM_PROVIDER_BASE_URL, MAIN_MODEL
pnpm run dev

Optionally build the search index for RAG:

pnpm run build-index

Open http://localhost:5173.

You'll need Node.js 26.x (see devEngines) and an OpenAI-compatible LLM endpoint. LM Studio, Ollama, or vLLM for local. Any cloud provider works too. Docker is optional but recommended for deployment.

MCP_SERVERS is a JSON array with $VAR substitution support:

[
  {
    "id": "github",
    "label": "GitHub",
    "url": "ttps://api.githubcopilot.com/mcp/",
    "type": "remote",
    "token": "$GITHUB_TOKEN"
  },
  {
    "id": "macula",
    "label": "Macula",
    "url": "https://u.macula.link/mcp",
    "type": "remote",
    "token": "public"
  }
]

Content lives in markdown files. Here's where everything is:

After changing content, rebuild the search index:

pnpm run build-index

Blog posts support GitHub-style admonition callouts. Use them inside blockquotes to highlight important information:

> [!INFO]
> Your file will be processed within 24 hours.

> [!WARNING]
> This operation cannot be undone.

> [!ERROR]
> Connection failed. Check your credentials.

> [!SUCCESS]
> Migration completed successfully.

Renders as color-coded callout boxes with left border accent and tinted background.

Supported types: INFO, WARNING, ERROR, SUCCESS.

Implemented via custom rehype plugin (src/lib/server/rehype-admonitions.ts) with no extra dependencies.

# Build and run
docker compose up --build -d

# Or build manually
docker build -t woss/woss-io .
docker run -p 3000:3000 --env-file .env -v ./data:/app/data woss/woss-io

Docker Compose stacks:

• woss: Main SvelteKit app (port 5173 → 3000)
• init: One-shot search index builder
• zinalog: Log aggregation dashboard (port 4000)

The tech side in one shot:

• Framework: SvelteKit 2, Svelte 5 (runes)
• Styling: Tailwind CSS v4, Tailwind Typography, sv5ui
• Runtime: Node.js 26, pnpm
• Database: SQLite (better-sqlite3)
• Vector Index: USearch (ANN)
• AI SDK: Vercel AI SDK (streamText, tool calling)
• Embeddings: HuggingFace Transformers.js (ONNX), in-process, no external API
• MCP Client: @modelcontextprotocol/sdk
• OG Images: Satori + resvg-js
• Logging: LogTape + ZinaLog
• Container: Docker + Docker Compose

AGPL-3.0-only