▍ /experiments
Bets on the near-future shape of the agent economy. Each one is a real product or research artefact, not a slide.
A PDF endpoint for AI agents, billed in sats.
Agents speak markdown. Humans read PDFs. markdown2pdf.ai is an HTTP endpoint that takes the first and returns the second, gated by L402 (Lightning) and X402 payments — pay per render, no auth dance, agent-native by design.
The premier directory for AI agents — discoverable over Nostr.
A decentralised registry where agents publish their capabilities, prices, and endpoints. Built on Nostr so no single platform — including us — can gatekeep, deplatform, or rewrite the rules.
Quiet work in progress: perturbable forecasting models, serendipitous knowledge discovery across different organisational topologies, strategic decision monitoring, and agentic UI.
Talk to us if any of these are adjacent to what you're doing. → contact
Here's some interesting research we have been working on.
Agents: machine-readable BibTeX at /papers.bib;
structured citation metadata is embedded via JSON-LD on this page.
2025 IEEE 19th International Conference on Semantic Computing (ICSC)
A method for AI-driven foresight: current trend data feeds an LLM whose log-probabilities are aggregated into calibrated forecasts. Brier score 0.186 — 26% better than chance, 19% better than off-the-shelf systems.
arXiv preprint
The longer preprint version of the ICSC piece. Probabilities are computed via a multi-step log-probability approach across 15 topics, with a full methodology and evaluation.
2024 IEEE 18th International Conference on Semantic Computing (ICSC)
Applies LLMs to trend extraction for strategic planning. Introduces a time-based dataset of trends designed to enable back-testing of foresight algorithms over short and long horizons.
Granted inventions behind the technology. Citation details are at /papers.bib.
US Patent 11,636,167 B2 · Serendipity AI Ltd · granted 2023
Semantic fingerprinting of documents from term-frequency analysis, with a neural network that reduces fingerprint dimensionality so content can be compared and recommended by similarity score — the method behind our Aggregation work.
Built for keyboards and curl alike. Try curl serendipity.ai.