talk to us

language models translate.
solvers decide.

kint uses each technology where it is strongest. Language models turn the planners' words into a model. Mathematical solvers find the best plan and report its gap. Machine learning covers what only your data knows.

show us one decisionhow it works
Foundations · technology
words · model · planthe solver has the last word
Who does what in a run

every step has one owner.

A decision passes through people, a language model, a solver and the platform. The record keeps their parts apart.

  1. 01 · your planners

    describe

    Goal and rules in the planners' own words, plus the data they use today.

  2. 02 · language model

    draft the model

    kint's decision agent reads the description and the data and writes variables, rules and goal.

  3. 03 · your experts · kint

    review

    In the pilot, kint's forward deployed engineers review the model with domain experts, in plain language.

  4. 04 · solver

    solve

    A mathematical solver searches the plans and returns the best one with its bound and gap.

  5. 05 · platform

    record

    Data version, model version, settings and releases, stored so the plan can be re-checked.

the language model never computes the plan.

It translates, and people check the translation. The plan comes from a solver. Each run pins the data version, the model version and the solver settings, so the plan can be re-checked against the record.

Mathematical optimization

established methods, chosen per problem.

kint picks the formulation from the structure of the problem. One modelling layer sits over several solvers, so the same model runs on open-source or commercial engines.

Optimization methods by problem structure
MethodStructureHow it is solved
LPContinuous quantities, linear rulesSimplex and interior-point methods. Allocation, blending, flows.
MIPYes-or-no and whole-number choicesBranch and bound with cutting planes. Shifts, routes, sites.
CPLogic, sequences, time windowsConstraint propagation and search. Timetables, job shops.
QPQuadratic goal, linear rulesInterior-point and active-set methods. Portfolios, load balancing.
Solvers

open and commercial

  • LP and MIP solversopen source, included
  • CP solveropen source, included
  • commercial MIP solverssupported
  • GPU serversconnected per engagement

Open-source solvers are included. Commercial solvers are supported; licences are arranged per engagement. kint matches the solver to the problem type. Your own GPU servers or cloud GPUs can be connected per engagement and run GPU-capable solvers. They pay off for very large linear programs and for routing. Many mixed-integer models solve just as fast on CPUs.

how GPU servers are connected
Machine learning

for what only the data knows.

Learned models enter where no formula exists, or where a forecast feeds the optimization. Each one reports how sure it is.

  • surrogate models

    Ensembles learn how outcomes respond to decisions.

    gradient-boosted trees · Gaussian processes

  • uncertainty

    Predictions come with an uncertainty interval.

    conformal prediction

  • search

    Single goals and trade-offs between several goals on the learned model.

    differential evolution · NSGA-II · NSGA-III

  • explanation

    Which inputs move the outcome, and how, for each candidate.

    per candidate

  • forecasts

    Demand, arrivals or prices as input to the plan. The plan states which forecast it used.

    time series

  • warm starts

    Learned predictions give the solver a good starting point. The bound and gap still come from the solver.

    solver hints

Architecture

built to sit inside other systems.

kint runs above your ERP, TOS, MES or VPP, or inside a vendor's product. Data comes in as read-only exports. Approved plans go back only through an API you control; the control system decides whether to apply them.

api-first

HTTP API with OpenAPI documentation, called from your systems.

REST · OpenAPI

live progress

Live progress of a solve over WebSocket.

WebSocket

durable workflows

Runs are versioned graphs of steps. After a crash or restart, the run continues at the step where it stopped.

workflow as code

re-check

From the pinned inputs, the plan can be re-checked against the record.

content-addressed inputs

oem-ready

Vendors embed the engine under their own brand and contract. Their customers see their product.

oem

where it runs

Hosted in Germany and operated by kint. For operations that must not leave your network, we scope an installation inside it.

security

see the stack on your problem.

Bring one decision. We show you the model the agent drafts, the solver kint picks and the record the run leaves behind.

start with one decision
Optimization
LP · MIP · CP · QP
Learning
Surrogate ensembles with uncertainty intervals
Interfaces
REST API · WebSocket · OEM
Contact
hello@kint.pro