Michał
Zagalski

A crypto data site needed accurate, current coin descriptions in several languages, with no invented facts and at a known cost per text.

A resumable pipeline crawls each project's own website, a judge model decides whether the material is good enough, an open model on Cloudflare Workers AI writes only from those sources, a sentence-level grounding check deletes unsupported claims, and translations into 8 languages pass a number and glossary check. Drafts that fail go to a human review queue.

Five models compared on the same top-30 coins through AI Gateway before the top-100 run; cost per description is read from the gateway's analytics.

Let AI agents pay for market-data API calls per request, without accounts or API keys.

Built Cloudflare Workers that gate API routes behind x402 (USDC on Base) and the Machine Payments Protocol, with routes and pricing configured in D1, a shared request log and an admin dashboard.

An online door & window shop needed unique product descriptions, with no invented product facts.

The agent collects the manufacturer's catalog text, a local LLM (Qwen 3.5 on llama-server) writes only from that source, then gates run from cheapest to most expensive: rule checks for invented facts, a similarity gate for uniqueness, and an independent Gemini-based reviewer agent with a local fallback.

Failure mode is “no description”, never “wrong description”. Measured end-to-end runs on real catalog items (acceptance rate, tokens/s, latency) go to the owner as the basis for choosing models.

A domain marketplace needed its catalog categorized and valued without manual work.

Batch classification of domains with structured JSON output and confidence scores, an LLM client with a local primary model (Ollama) and OpenAI fallback, an AI valuation assistant and a Telegram bot for admin operations.

Local model first, paid API only as fallback: cost and availability under control.

Find duplicate and off-topic pages on a website without sending its content to an external API.

Local embeddings (bge-small via Transformers.js) stored in pgvector with an HNSW index, computed by BullMQ background workers next to 50+ technical SEO checks.

GPU rent is often the biggest line in an AI startup's budget and its price moves; a small team cannot buy a fixed-price hedge for H100 hours.

A cash-settled fixed-for-floating swap on an H100 GPU-hour index: an Anchor program on Solana devnet, a Cloudflare Worker that publishes the index on-chain and settles each period, and a web app for both sides. The swap is capped at ±30% of the fixed price and each side deposits that cap up front, so there are no margin calls and no liquidations.

Built in October 2026 for Colosseum Crypto World's Fair; runs on devnet with mock USDC.