Heterogeneous AI runtime
Not every task needs an LLM.
Reading text, predicting risk, forecasting demand, judging a situation and calculating an amount are different problems. Syrtis composes the right technologies for each step.
Your data already knows a lot.
Estimate risk, classify a record, predict a value from tabular data. A PFN works on your business data, alongside language models.
Local or cloud. Models change. Your rules stay.
Explore the architectureSpecialist models in one process.
A language model understands and writes. A PFN works with tabular data. A time-series model produces a forecast. A judgment model returns a typed choice or score. Code calculates and applies exact rules.
- LLM — Claude, GPT, Gemini, Mistral
- PFN — TabPFN / Prior Labs
- Forecasting — Chronos, TimesFM
- Judgment — Jev / TypeSafe AI
- Execution — code and business rules
Uncertainty is part of the result.
A prediction or judgment does not become certain because its output is structured. Your business logic sets action thresholds, checks and the handoff to human approval.
Change models without losing your bearings.
Test scenarios compare two versions on the same business situations. Changes to providers, instructions and models remain identifiable in the history.
A patent filing for the orchestration method is in preparation.
It starts with
one process.
The one your team does by hand, every week.
What if your software took care of it?