oddsys.org

High-page-count OCR: scanned PDF into clean text (Tesseract, English, machine-printed documents), up to 100 pages per call — four times the /ocr cap. Runs under a 240s wall-clock budget; slower documents return the completed pages with truncatedReason "time-budget". Text-layer PDFs are extracted the fast way at the same price.

Generate sentence embeddings (all-MiniLM-L6-v2, 384-dim, mean-pooled + L2-normalized) for up to 96 texts per call. Drop-in compatible with any all-MiniLM-L6-v2 index. Pairs with /chunk for a complete RAG pipeline.

Live foreign-exchange reference rates (European Central Bank daily fixes, ~30 major currencies): any cross-rate from any base, optional amount conversion. Refreshed daily; response carries the rate date.

Fetch a public web page and return clean markdown (or structured JSON) — URL in, agent-ready content out, typically 80-85% fewer tokens than raw HTML. Handles redirects and compression.

Modify an image in one call: crop, resize, convert to greyscale, and re-encode (PNG/JPEG/WebP) — applied in that order, any combination. Deterministic, no upscaling surprises.

Detect and locate objects in an image (YOLOX-nano, COCO classes): people, pets (dog/cat/horse), vehicles, and 70+ more. Returns per-class counts and pixel bounding boxes. Detection only — no identification. Best on clear, medium-to-large objects.

Validate and normalize phone numbers (Google libphonenumber): validity, E.164/international/national formats, country, region, and line type (mobile/fixed/VoIP). Single number or batches up to 100.

Generate a QR code from text as PNG (base64) or SVG. Custom module size, quiet-zone margin, and hex colours. Deterministic, capacity-capped at 2953 bytes. Set delivery to "url" to receive a short-lived download link instead of inline data.

Rerank candidate texts against a query with a cross-encoder (ms-marco-MiniLM-L-6-v2): joint query+document scoring, much sharper than cosine similarity over embeddings. Returns ranked indices with 0-1 relevance scores. Up to 64 texts per call. Pairs with /embed for a recall-then-precision RAG pipeline.

Fuzzy sanctions screening against the US OFAC SDN list (public-domain source, refreshed daily): name in, scored matches with programs and aliases out. Batches up to 50. Coverage is OFAC SDN only — stated in every response.

Decode EVM calldata or a Base-mainnet transaction into human-readable form: function selector, matched signature, named args, plus ERC-20/USDC transfer events from the receipt — verify an x402 settlement receipt (tx hash in, payer/payee/amount out) in one call.

Convert raw HTML into clean markdown. Typically cuts page tokens by 80-85% before they hit an LLM context window.

Extract clean markdown text from a PDF (text-layer PDFs; no OCR). Turns a document an agent can't cheaply read into compact tokens.

Convert data between JSON, YAML, CSV, and XML. Deterministic format juggling done off-context, with edge cases handled.

OCR a scanned PDF into clean text (Tesseract, English, machine-printed documents). Up to 25 pages per call. PDFs that turn out to have a text layer are extracted the fast way at the same price.

Split a document into token-bounded chunks at natural boundaries (paragraphs/sentences), with optional overlap. Ready-to-embed pieces for RAG, no tokens spent on the splitting logic.

Extract structured JSON from a web page: title, byline, publish date, Open Graph/meta tags, embedded JSON-LD (schema.org), links, and images. Deterministic, no LLM.