Case Studies

What it does

Dashrock is a serverless analytics platform for AWS Bedrock. It collects and visualizes LLM usage data across teams and applications — costs, token counts, model distribution, latency, and trends over time. Teams define apps and members in a simple YAML config, deploy via SST in minutes, and get a full analytics dashboard with no ongoing infrastructure to manage.

Tech Stack

How DAUB is used

Dashrock uses DAUB for the entire dashboard UI — cards, tables, navigation, layout containers, and form elements. The classless CSS approach let Marat build clean, consistent interfaces rapidly with AI code assistance. DAUB's theme variables provide dark/light mode and the tactile, considered aesthetic that makes the analytics feel polished rather than utilitarian.

DAUB is credited in the project footer ("UI/UX built with daub.dev") and listed as a key dependency in the README tech stack.

Key Features

• 7 dashboard views: Overview, Applications, Models, Teams, Token Usage, Usage Trends, Cost Analysis
• YAML-based team and application configuration
• CSV export for all data tables
• Date range filtering with presets
• Custom domain support via CloudFront
• HTTP Basic Auth for access control
• ~3 minute deploy via SST

"Daub gives you the speed of a utility framework with the polish of a design system. Built two production apps with it and would reach for it again without hesitation."

Marat, Dashrock

What it does

Tamp is a local HTTP proxy that intercepts Claude API requests and applies a multi-stage compression pipeline before forwarding them upstream. Claude Code (and any Anthropic API client) sends verbose tool results on every turn — pretty-printed JSON from Read, raw CLI output from Bash, full file listings from Grep and Glob. These accumulate to 100K+ tokens in minutes, driving up cost and exhausting context windows.

Tamp fixes this transparently. Set ANTHROPIC_BASE_URL=http://localhost:7778, run npx @sliday/tamp, and every API call gets compressed on the fly. Responses stream back untouched. Claude doesn't know the difference.

3-Stage Compression Pipeline

Tamp classifies each content block (JSON, prose, already-compressed) and routes it through the appropriate stages:

• Stage 1 — JSON Minification: Pretty-printed JSON is minified. Zero risk — Claude handles minified JSON perfectly. ~15-20% savings.
• Stage 2 — TOON Encoding: Tabular/array JSON is converted to TOON format, a compact alternative representation. ~40-60% savings over minified JSON.
• Stage 3 — LLMLingua-2 (optional): Natural language prose is compressed via a Python sidecar. Lossy but meaning-preserving. 2-20x compression.

Images, system blocks, tool schemas, and content under 200 characters are never touched. A cache-safe mode (default) only compresses new content in the latest user message, preserving Anthropic prompt cache hit rates.

Tech Stack

How DAUB is used

The Tamp marketing site (tamp.dev) is built entirely with DAUB's classless CSS. The landing page uses DAUB's typography system, spacing tokens, and theme variables to create a clean developer-focused presentation — no custom CSS classes needed. DAUB's CDN delivery (daub-ui@latest) means the site loads fast with zero build step.

The tactile aesthetic works particularly well for developer tools — it makes a terminal proxy feel polished rather than bare-bones, which matters when you're asking developers to route their API traffic through your software.

Key Features

• One-command install: npx @sliday/tamp
• Per-block compression stats logged in real-time
• Session-wide savings tracking (cumulative chars saved)
• Selective compression — only POST /v1/messages, passthrough everything else
• Cache-safe mode preserves Anthropic prompt cache hits
• Graceful error recovery — compression failures fall back to original content
• Optional LLMLingua-2 sidecar for prose compression
• 570 LOC total — small, auditable, no framework dependencies

Why it matters

Extended Claude Code sessions with heavy tool use can burn through context windows and rack up significant API costs. Tamp addresses this at the transport layer — no changes to your workflow, no changes to Claude's behavior, just smaller requests. The 50% savings compound across a full development session: a typical hour of Claude Code work that would consume 500K input tokens drops to ~250K. That's real money at Anthropic's pricing.

What it does

Dwarf Land is a persistent civilization simulator rendered as an emoji tile planet — continents, oceans, mountains, jungles, deserts, and a map that wraps horizontally like a globe. Dwarves with D&D stats and culturally accurate names mine, farm, craft, trade, raise families, tame animals, and build roads between cities. Nobody scripts their behavior: each decision comes from an LLM call routed through a 4-tier model system, and the world keeps running whether you watch or not. State saves every 60 seconds; your dwarves remember everything.

4-Tier AI Brains

Every dwarf thinks through OpenRouter via the Vercel AI SDK. Model routing depends on the decision's complexity — and on sponsorship:

• Simple — basic needs: eat, sleep, mine. The free default.
• Medium — social behavior and planning. Bronze sponsorship ($1, 100 calls).
• Complex — strategy and relationships. Silver ($3, 75 calls).
• Premium — deep reasoning with full personality context. Gold ($10, 100 calls).

AI calls are fire-and-forget: the game never blocks on a response. An intent cache stores each decision until the dwarf is idle enough to execute it, and hard budget caps ($8.50/hr total) keep the world from thinking itself bankrupt. Sponsoring a dwarf's brain goes through Polar.sh checkout — the dwarf gets a star badge and visibly smarter behavior.

Tech Stack

How DAUB is used

The entire game chrome is DAUB in the grunge theme (data-theme="grunge") — dark, tactile HUD panels that let the emoji world own the screen. The welcome and How-it-works overlays are db-modal with db-prose content, sponsorship pricing renders as a compact db-table, and buttons, scroll areas, and toasts come straight from the component library — driven at runtime by DAUB.openModal() and DAUB.toast().

The whole frontend is a single HTML file pulling [email protected] from CDN. No build step, no framework — vanilla JS canvas for the world, DAUB for everything with text in it.

Key Features

• Up to 300 autonomous dwarves with D&D stats, aging, families, and AI-generated backstories
• 125+ real-world cities across 7 continents with culturally named populations
• 4-tier road system (dirt → gravel → asphalt → railroad) built, upgraded, and repaired by dwarves
• Infinite crafting: combine any two items — unknown combos resolved by AI and cached forever
• 24 wildlife species with taming, hunting, and combat encounters
• Horse carts, cars, and trains; autonomous freight balances resources between cities
• Live sponsorship via Polar.sh upgrades individual dwarves' AI reasoning
• 453 Vitest tests; world state autosaves every 60 seconds

Why it matters

Most DAUB deployments are dashboards and marketing sites. Dwarf Land stretches the library into game HUD territory — modals, tables, and toasts floating over a 60fps canvas — and the grunge theme handles it without a line of custom component CSS. It also demonstrates LLM calls as a game mechanic with real economics: budget caps on the house's spend, micro-payments to raise a single agent's intelligence.