Live demos · real database underneath

See what one database
collapses for you.

Six auto-playing demos. Each one is a 60-second narrative driven by an actual SynapCores AIDB instance — SQL, Cypher, vector similarity, and AutoML in one connection. Hit Play full demo for the recording, or jump to the source.

LIVE
AI Operations · MCP

Detective

Claude opens the database and finds the culprit. 8 MCP tools, 1 Cypher walk, 47 seconds — Northwind's lost revenue traced through one vendor.

MCP × 8 toolsCypher MATCHVector + AutoML
Get the code
LIVE
Cybersecurity

SentryGraph

Phish → credential stuffing → account takeover → BEC, in 60 seconds. JA3 botnet revealed by one Cypher walk.

HTTP + auth streamCypher 1-hopincident_id stitch
Get the code
LIVE
AI Safety

GuardGraph

A GraphRAG-powered firewall for LLM outputs — catches PII leaks, policy drift, hallucinations, and coordinated abuse rings.

EMBED()COSINE_SIMILARITYCypher incident graph
Get the code
LIVE
Finance

TradeGraph

Wash-trade and shell-company cycle detection over a live trade graph. Vector similarity surfaces the look-alike actors.

Cypher cyclesCOSINE_SIMILARITYAutoML risk score
Get the code
LIVE
Pharma

RxGraph

Drug repurposing across mechanism, target, and trial data — one query stitches the graph and the embedding space.

Vector + Graph joinCypher 2-hopNL2SQL
Get the code
LIVE
Fraud & Risk

RiskGraph

Fraud-ring detection on a real-time transaction graph, with behavioural embeddings and per-edge risk scoring.

Cypher ringsBehaviour embeddingsAutoML
Get the code
LIVE
Sports & Culture

World Cup Scout

The legend's modern echo, the draw as a graph, ten thousand tournaments simulated. One SQL surface.

EMBED()COSINE_SIMILARITYCypher rivalries
Honest numbers · measured, not modeled

How fast does in-database AutoML actually train?

Same SynapCores CE binary as the demos above. Same CREATE EXPERIMENT SQL call. Binary classification, 3 trials, 5-fold CV, all CPU — no GPU, no remote inference.

Training time vs. row count
Rows trainedColdWarmBest score
2000.51 s0.03 s0.95
2,0000.67 s0.18 s0.97
10,0001.27 s0.84 s0.97
100,00011.21 s10.62 s0.97
Measured 2026-05-16 against fix/v1.6.4-demo-bugs HEAD 61ae01f8. Binary classification on a synthetic loyalty-churn dataset with clear signal. Times include SQL parse, plan, execute, fold splits, model fits, validation, and Platt calibration. Cold = first call after gateway start. Warm = subsequent calls.
Hardware
CPU
Intel Core i5-10400F
6 cores / 12 threads · 2.9 GHz
RAM
32 GB
GPU
RTX 3060
present, unused — community-edition build is CPU-only
OS
Ubuntu 22.04.5 LTS
Inference
Pure Rust + rayon
12 worker threads

100,000 rows trained in 11 seconds on a $400 CPU — and that's the cold-cache first-call number. Warm calls are sub-second up to 10k. GPU acceleration is a build flag away (gated off in the community edition by default); turning it on for a workload that already fits CPU is genuinely optional.

Ready to write your own?

187 ready-to-run recipes — GraphRAG, fraud rings, AutoML, multimodal queries. Copy any one into the bundled web UI and watch it run.

Browse recipes →