keyronne.com

Exact computation LLMs cannot do reliably: evaluate arithmetic expressions (trig, logs, powers, factorials, constants), convert between 80+ units across 11 dimensions (length, mass, temperature, data, speed, energy, pressure...), and compute descriptive statistics (mean, median, stdev, percentiles) over up to 100k numbers. Deterministic, no hallucinated arithmetic.

Repair malformed JSON text into valid parsed JSON: markdown code fences, single quotes, smart quotes, unquoted keys, trailing commas, comments, Python literals (True/None), raw newlines in strings, and truncated output (auto-closes strings, arrays and objects). Returns the parsed value plus a log of every repair applied. Built for LLM output that almost parses.

Validate a JSON value against a JSON Schema and coerce common LLM-output mistakes (numbers as strings, stringified booleans, scalars where arrays are expected, missing defaults, extra properties). Returns the coerced value plus a structured error report. Built for cleaning up LLM tool-call arguments and structured outputs.

Split long text into retrieval-ready chunks along natural boundaries — paragraphs, sentences, or markdown sections — with configurable size, overlap, character offsets and approximate token counts per chunk. Deterministic and language-agnostic. Built for RAG ingestion and embedding pipelines.

Compare two texts with a Myers diff over lines, words, or characters. Returns structured edit operations, a standard unified diff (line mode), and similarity statistics. Built for agents tracking document revisions, verifying edits, or reviewing generated changes.

Validate email addresses: RFC-style syntax checks, IDN domain normalization, live MX lookups with A/AAAA fallback (DNS-over-HTTPS), disposable-domain detection, role-address detection (info@, support@), free-provider flagging, and typo suggestions for popular providers (gmial.com -> gmail.com). Single address or batches up to 50.

Compute pairwise text similarity, find near-duplicates, and cluster a batch of texts (2–300 items). Character n-gram cosine similarity: language-agnostic, robust to small edits, no model calls. Returns similar pairs above a threshold, connected-component clusters, and the full matrix for small batches.

Current conditions plus hourly and daily forecasts for anywhere on Earth. Pass a major city name or lat/lon coordinates; get temperature, wind, humidity, pressure, cloud cover, condition codes and precipitation, with up to 90 hourly and 10 daily entries. Data from MET Norway (CC BY 4.0). No API keys, one call.

Convert a web page (by URL) or raw HTML into clean, readable Markdown plus structured metadata (title, description, Open Graph tags, canonical URL, language). Strips scripts, styles and boilerplate tags; preserves headings, links, images, lists, code blocks and tables. Built for agents that need web content as LLM-ready text.