Business data as geometry that AI agents navigate.
hypertopos turns relational data into a shared geometric space where:
- entities become positions, derived from their relationships and the population around them
- relationships become structure
- behavior becomes movement
- anomalies become distance from the population — and the direction of that distance explains them
It complements SQL: the database answers which rows match; hypertopos answers which entities are out of place, how that is changing, and why.
This layer sits between raw data and downstream systems:
- for exploration (analysts, agents)
- for feature generation (ML pipelines)
- for monitoring (drift, regime change)
No model training required.
Core engine. Builds a sphere from a YAML description and stores it as Lance datasets. Every session reads one consistent snapshot while builds, ingests and recalibrations continue underneath; what-if work — a recalibration without suspected influencers, a trial ingest — runs on a copy-free branch that is compared with main, then promoted or deleted.
Agent interface. Exposes the sphere as MCP tools — AI agents navigate, inspect, and reason about data structure directly. Runs over stdio or local HTTP, with an independent session per connected agent; responses are compact JSON built for an agent's context window.
Investigation workflows. Structured behaviors that guide agents through real tasks: anomaly triage, root-cause tracing, chain and fraud investigation, drift monitoring, sphere design.
Data → geometry (hypertopos-py) → tools (hypertopos-mcp) → reasoning (hypertopos-skills)
Research-stage project. Working code, reproducible benchmarks, active development. API may change.
pip install hypertopos
