See what your data
actually looks like.
Pulsar is a topological data analysis engine that finds the real structure in high-dimensional data — not the shape your model assumes. Instead of betting on one clustering or embedding, it sweeps hundreds and keeps only the structure that survives all of them.
Stop picking the pretty embedding.
K-means, UMAP, DBSCAN — each forces your data into a shape it picked, then ships the prettiest result. Nudge one parameter and the clusters dissolve. Pulsar takes the opposite bet: it trusts only the structure that holds no matter how you look.
“I see three groups. (Change k to six and none of this means anything anymore.)”
“Here are the communities the data actually has. Ask for more resolution — the structure holds.”
Judging your data from one clustering is like judging a sculpture from one photo — the angle alone can make it look like a blob, a cross, or a face. Pulsar photographs it from hundreds of angles simultaneously (the multiverse sweep), keeps only what shows up in every shot, and fuses them into one reliable picture: the Cosmic Graph.
Build hundreds of representations across a grid of parameters — photograph the data from every angle, not one.
Construct a graph for each. Any single one can be lucky or misleading — so none gets to decide alone.
Accumulate the evidence into one Cosmic Graph. Only structure that persists across every view survives.
Structure that persists across 100 representations is structure you can trust.
It's built on the THEMA algorithm — peer-reviewed in Nature Energy.
Pulsar recovered a landmark cardiology finding in 5 minutes.
Pulsar runs a full topological analysis of a public Myocardial Infarction dataset — live in the Gemini CLI through its open MCP integration. No labels, no clinical priors, no AHA guidelines.
The nitrate contraindication in right-ventricular MI took decades and a 40-patient study to codify. Pulsar recovered it from raw, unlabeled data in a single session.
No labels, no clinical knowledge — just the geometry of a public dataset, mapped from scratch and handed to the model to interpret.
It isolated a distinct cohort of seemingly low-risk patients who deteriorate catastrophically — a preload-dependent group that nitrates can tip into cardiogenic shock.
Runs entirely on the open-source engine and its MCP integration — nothing proprietary. Get it on GitHub→
Reproducible results, across domains.
Every output is grounded in structure that survived hundreds of modeling choices — not a single lucky run. The same engine holds up from clinical data to benchmark evals to energy policy.
faster than pure Python (Rust core)
for a full 4,000-map sweep
real regions vs. labels found in MMLU
coal plants sorted into retirement archetypes
Recover phenotypes from raw EHR data
Cluster patient trajectories — not just snapshots — to surface cohorts heading toward deterioration before their vitals visibly diverge.
Read the Pulsar field note→Map the real structure of your benchmarks
Expose model-specific failure zones invisible to leaderboard averages, and cover the same evaluation space with a fraction of the data.
Read the MMLU study→Find the archetypes inside complex systems
Reveal the natural groupings and hidden vulnerabilities embedded in operational data — across any high-dimensional dataset, no prior labels needed.
Read the Nature Energy study→Stop picking the pretty embedding. Start trusting the geometry.
Pulsar is free, open source, and production-ready. Pick your door.
Book a demo
See Pulsar map your dataset, live.
Open→View on GitHub
Inspect and run the engine. BSD-3, open source.
Open→Read the docs
Go deep on the method and the API.
Open→