
Chemistry
From a molecule name to its quantum circuit
Molecules come from PubChem and crystals from the Crystallography Open Database, drawn in 3D. A molecule can go straight to a quantum circuit plan for its ground-state energy.
SynapseX AI
From catalysts to orbits: agents that fetch real data, run the calculation and show you every number they computed, with a person in charge.
Capabilities

Agents
Use cases
Each card names the SynapseX tool that does the work. The tool fetches or computes the numbers, and the agent explains them.

Chemistry
Molecules come from PubChem and crystals from the Crystallography Open Database, drawn in 3D. A molecule can go straight to a quantum circuit plan for its ground-state energy.

Biology
Structures from the PDB and AlphaFold DB with measured contacts and confidence, records from UniProt, Ensembl, Reactome, ClinVar, GEO and PubMed, and genome tracks in a browser view.

Space
Where a planet, moon or spacecraft was, is or will be, read from JPL Horizons and drawn in 3D at real scale, next to an orbit you integrate yourself.

Rocket science
Dynamical systems integrated and plotted, with air temperature, pressure and density up to 86 km from the 1976 U.S. Standard Atmosphere.

Modern physics
Crystal structures, surfaces and vector fields, Schrödinger wave packets stepped live on a lattice, and surface and repetition codes with their decoding.

Natural science
Maps and choropleths, scientific data files such as VTK, coordinate tables and XYZ, and embedding spaces, with Python in an isolated sandbox for the analysis.
AI agents in catalysis research
Much of a catalysis study is routine: finding structures, setting up each calculation, launching runs, collecting the numbers and comparing them. Steps a researcher strings together by hand, often over days, an agent sets up in one conversation, and every number it reports was computed.

Step 1 of 4 · Find
Ground-state energies are implemented today. Reaction barriers, transition states and surface models are not.
Agents take on the routine of a computational study: structures, set-up, runs and the collection of results.
Upload data or index your papers in a Library, then let an agent explore it with Python in a sandbox and plot what it finds.
Agents that propose the next experiment and hand it to lab automation. This is not in the product today.
What SynapseX brings
Distances, atom counts, confidence scores and energies come from code. The model is told never to make one up in prose.
PubChem, the Crystallography Open Database, the PDB, AlphaFold DB, UniProt, Ensembl, Reactome, ClinVar, GEO, PubMed and JPL Horizons, each within its published rate limits.
A molecule becomes a circuit plan and runs on a free CPU simulator through QCOS. Execution on quantum hardware is switched off.
A task can fan out to parallel agents, combine their results and vote on them in a verify stage that fails closed.
Experiment plans, versioned datasets, run comparison and evidence receipts in SynapseX Lab.
Scheduled tasks with a tool allowlist and a time limit per run, suspended after repeated failures.
How an agent works
Agents name a connector and an operation. Tokens stay behind a gateway, and secrets are redacted.
Tool output is framed as untrusted, and probing is audited and refused. This is layered defence, not formal taint tracking.
Actions are classed as read, write, publish, financial or destructive, and publish and financial actions need approval.
Model programme
SoftQuantus is specialising models for quantum and scientific work with SFT and DPO. Until a programme model passes its evaluation, SynapseX runs models chosen for each kind of work.
SoftQuantus holds a EuroHPC AI Factory allocation (EHPC-AIF-2025LS13-019) for training compute.
A programme model serves only after it passes its evaluation. None has passed yet, so none serves today.
We will publish the size and results of the largest programme model when it serves, not before.
Where you use it
Multi-model chat, voice, documents and agents on the web.
Experiment plans, versioned datasets, run comparison and evidence receipts.
A desktop app with per-task git worktrees, and iOS and Android apps.
A terminal agent with a headless server and JavaScript SDK, and a REST API.
Vision
None of these exists in the product today.
Agents that plan the next experiment from the last result and hand it to lab automation.
A model running only for your organisation, inside its own boundary.
Agent machines in hardware enclaves, with keys you hold.
Observers beside every turn that check the work, recall memory and track the goal.
Once a process reads untrusted content, its network access waits for a person.
Trust
Film and photographs made for this page by SoftQuantus. They illustrate each field and do not show SoftQuantus hardware, staff, customers or results.