SynapCores / Playground
Run SQL, Cypher and in-database ML in your browser
A live SynapCores engine loaded with one small dataset: 123 card payments from 41 accounts, a fraud model trained on them in SQL, and the graph of which accounts share devices and cards, plus a few support notes stored with vector embeddings. Run the steps below in order. They show what per-row ML catches, what it misses, and how one graph query fills the gap, then search the case notes by meaning. Every box is editable; your first few runs need no signup.
Step 01 · SQL
Score every payment with ML
An isolation-forest model (trained in SQL with AutoML) scores all 123 payments inside the SELECT. ACC-33’s three 3am cross-border payments float to the top at ~0.81; everything else sits near 0.5.
Step 02 · SQL
The model can’t see the ring
Five accounts (ACC-11 to ACC-15) are a fraud ring. Each payment looks ordinary on its own, so the same model scores all 15 of them below 0.6. Per-row ML has nothing to flag.
Step 03 · Cypher
Walk the graph from one chargeback
Start at ACC-13 (one chargeback) and follow shared devices and cards 2–4 hops. Cypher finds ACC-14 and ACC-12, then ACC-15 and ACC-11. The household pair ACC-21/ACC-22 shares a device too, but isn’t connected to the ring, so it isn’t returned.
Step 04 · SQL
One review queue: ML + graph
Combine both signals in one query: accounts the model flags (score > 0.7) plus accounts the graph linked to the chargeback. Result: six accounts, ACC-33 from the model and the five ring members from the graph.
Step 05 · SQL
Search case notes by meaning
Support notes are stored with a VECTOR(384) embedding column. EMBED turns your question into a vector inside the query and COSINE_SIMILARITY ranks the notes by meaning, not keywords. The three account-takeover notes from the ring (ACC-12, ACC-14, ACC-11) come first: a new phone, a shared device fingerprint and a password reset from an unfamiliar device. Billing and delivery complaints rank last.
What’s not here
This sandbox has no LLM. It’s read-only, and queries are capped at 100 rows and 5 seconds. Natural-language SQL, GENERATE, AGENT_RUN, agent memory and RAG all need a model, so they run on your own install. Community Edition is free and runs those features with a local model, and nothing leaves your machine.
Run it on your own data
Install SynapCores Community Edition, load the full recipe, and swap in your own payments and identity graph.