Send JSON. Get results. No CUDA, no Python, no GPU drivers.
Euriklis is a computational platform that exposes production-grade numerical algorithms through a single language-agnostic JSON API. Whether your stack is Java, PHP, R, Excel, a browser, or an AI agent — you describe what to compute and we return the result.
Why this exists
Modern quantitative work — portfolio optimisation, risk modelling, regression, hypothesis testing — depends on numerical routines that are either locked inside Python notebooks or buried in proprietary enterprise software. Wiring them into a Java backend, a PHP web app, or an AI agent's tool chain is friction we shouldn't be paying in 2026.
We remove the friction. Computation becomes one HTTP call.
What's live today
Production-grade CPU compute. Every operation runs on a multi-worker parallel runtime built on SharedArrayBuffer and Atomics, calibrated per workload from empirical benchmarks. Matrix multiplication scales across cores; Cholesky, LU and Householder QR dispatch to sequential, Atomics-coordinated parallel, or blocked WY pipelines depending on problem size — the fast path is always selected automatically.
High-precision numerics. IEEE 754 float64 by default. Partial pivoting in LU and Gauss-Jordan. Dtype-aware singularity tolerance. SPD detection in Cholesky. No silent precision degradation, no opaque heuristics.
Graph-based execution. A single request can describe an entire pipeline as a directed acyclic graph — covariance → Cholesky → solve → weights → CVaR — submitted, parallelised, and returned in one round trip.
Built-in financial and statistical primitives.
- Markowitz portfolio optimisation (min-variance, max-Sharpe)
- Conditional Value-at-Risk (CVaR) tail-risk analysis
- Linear regression with diagnostic suite (residuals, R², heteroscedasticity)
- Matrix factorisations: Cholesky LLᵀ, LU + partial pivoting, Householder QR
- Statistical hypothesis testing
Validated in production. The mathematics layer underpins live commercial systems — a pharmacy-network graph-matching engine for Novo Nordisk Bulgaria (8 287 entities matched) and the Cocosolis AI customer-support orchestrator.
What's coming
The same JSON API will route to native GPU execution across NVIDIA, AMD and Intel through our in-house TypeScript-to-LLVM-IR compiler, currently in proof-of-concept with measured performance at 100.8 % – 108 % of hand-written CUDA C on validated vector workloads. The migration is invisible to clients: existing API calls automatically accelerate when the hardware path lights up.
Model Context Protocol (MCP) integration follows, making every endpoint directly callable by AI agents without human glue code.
Sign up, generate an API key, send your first request in under five minutes. Free tier covers exploratory workloads; paid tiers scale with throughput.