Theovya Optimization Engine

The LLM models. The solver solves.

Translates business goals and constraints into formal optimization problems, solved by exact methods where optimality is provable and heuristics where scale demands it.

Status

on the roadmap

Answers

what should we do

Builds on

Prediction Engine forecasts

Solvers

exact methods and heuristics

How it works

Business problem in, solver-grade decision out.

Given these forecasts and my warehouse capacities, what is the cheapest replenishment plan?

  1. 01The same proprietary agentic structure as prediction, applied to prescriptive questions, starting from forecasts the Prediction Engine already produced.
  2. 02The agent translates the business problem into a formal model: decision variables, constraints, and an objective.
  3. 03Exact methods (linear and mixed-integer programming solvers) run where provable optimality is achievable.
  4. 04Heuristics and metaheuristics take over where problem scale demands fast, high-quality approximate solutions.

Why the answer is provable, not plausible

An optimization answer is never text a language model composed. It is the solution vector a solver returned. The model’s role ends at formulation; the mathematics runs on algorithms with decades of industrial validation behind them.

That division of labour carries a safety property: a mis-formulated model produces infeasible or visibly wrong output, not a plausible-sounding fabrication. Constraints are checkable, objectives computable, and exact methods return a provable optimality guarantee.

Validated by current research

OptiMUS (Stanford) showed agent-based LLM workflows reliably translating natural-language problem descriptions into MILP formulations solved by industrial solvers. OptiTree (NeurIPS 2025) added hierarchical tree search that decomposes complex OR problems into subproblems, improving modelling accuracy by over 10% on hard benchmarks.

The three engines

Three engines. Independent, composable, built to compound.

Each engine answers the next question on the maturity ladder. A business runs one, some, or all three, configured to what it actually needs, not a fixed bundle.

01 Available today

Answers what happened

Analytics Engine

Turns plain-language questions into live queries, charts, dashboards, and scheduled alerts, grounded in the company’s real schema, with the exact query behind every answer.

Explore the Analytics Engine
02 In progress

Answers what will happen

Prediction Engine

Builds, trains, and validates real machine-learning models from the user’s own data, guided entirely by prompts, and returns honest held-out metrics with every model.

Explore the Prediction Engine
03 On the roadmap

Answers what should we do

Optimization Engine

Translates business goals and constraints into formal optimization problems, solved by exact methods where optimality is provable and heuristics where scale demands it.

You are here

All three share one design principle: the language model never invents an answer. It only translates intent into verifiable computation.

Use cases

The Optimization Engine, applied.

Each case ends here: what happened and what will happen become a constrained, actionable decision.

08 cases

AnalyticsPredictionOptimization

  4. 07

    Medical servicesClaraCare

    Booking the empty chair

    Filling the empty chair without filling the waiting room.

    Read the case

It is the most valuable question, and the one least often answered properly.

Once a business knows what happened and what will happen, one question remains: what should we do? Answering it is operations research - turning goals, limits and trade-offs into a mathematical model and solving it - and specialists who can do that are among the scarcest people in industry. So carrier splits, rosters, order quantities, offers and budgets are set by rule of thumb, and nobody learns how far from the best they are.

The Optimization Engine takes the question in plain language and writes the formal model - what to decide, what must hold, what to minimize or maximize - in terms the person asking can check. A solver does the mathematics: exact methods where optimality can be proven, heuristics with a stated quality gap where scale demands them. If the requirements contradict each other, it says so and names them.

The language model is the interface. Solvers are the engine room.

The Optimization Engine turns Prediction Engine output into constrained, actionable decisions. Start with the Analytics Engine, live today and connected read-only to systems already in place.

Request a conversation

Theovya Optimization Engine

The LLM models. The solver solves.

Translates business goals and constraints into formal optimization problems, solved by exact methods where optimality is provable and heuristics where scale demands it.

REQUEST A CONVERSATION See how it works

Status

on the roadmap

Answers

what should we do

Builds on

Prediction Engine forecasts

Solvers

exact methods and heuristics

How it works

Business problem in, solver-grade decision out.

Given these forecasts and my warehouse capacities, what is the cheapest replenishment plan?

  1. 01The same proprietary agentic structure as prediction, applied to prescriptive questions, starting from forecasts the Prediction Engine already produced.
  2. 02The agent translates the business problem into a formal model: decision variables, constraints, and an objective.
  3. 03Exact methods (linear and mixed-integer programming solvers) run where provable optimality is achievable.
  4. 04Heuristics and metaheuristics take over where problem scale demands fast, high-quality approximate solutions.

Why the answer is provable, not plausible

An optimization answer is never text a language model composed. It is the solution vector a solver returned. The model’s role ends at formulation; the mathematics runs on algorithms with decades of industrial validation behind them.

That division of labour carries a safety property: a mis-formulated model produces infeasible or visibly wrong output, not a plausible-sounding fabrication. Constraints are checkable, objectives computable, and exact methods return a provable optimality guarantee.

Validated by current research

OptiMUS (Stanford) showed agent-based LLM workflows reliably translating natural-language problem descriptions into MILP formulations solved by industrial solvers. OptiTree (NeurIPS 2025) added hierarchical tree search that decomposes complex OR problems into subproblems, improving modelling accuracy by over 10% on hard benchmarks.

The three engines

Three engines. Independent, composable, built to compound.

Each engine answers the next question on the maturity ladder. A business runs one, some, or all three, configured to what it actually needs, not a fixed bundle.

01 Available today

Answers what happened

Analytics Engine

Turns plain-language questions into live queries, charts, dashboards, and scheduled alerts, grounded in the company’s real schema, with the exact query behind every answer.

Explore the Analytics Engine
02 In progress

Answers what will happen

Prediction Engine

Builds, trains, and validates real machine-learning models from the user’s own data, guided entirely by prompts, and returns honest held-out metrics with every model.

Explore the Prediction Engine
03 On the roadmap

Answers what should we do

Optimization Engine

Translates business goals and constraints into formal optimization problems, solved by exact methods where optimality is provable and heuristics where scale demands it.

You are here

All three share one design principle: the language model never invents an answer. It only translates intent into verifiable computation.

Use cases

The Optimization Engine, applied.

Each case ends here: what happened and what will happen become a constrained, actionable decision.

08 cases

AnalyticsPredictionOptimization

  4. 07

    Medical servicesClaraCare

    Booking the empty chair

    Filling the empty chair without filling the waiting room.

    Read the case

It is the most valuable question, and the one least often answered properly.

Once a business knows what happened and what will happen, one question remains: what should we do? Answering it is operations research - turning goals, limits and trade-offs into a mathematical model and solving it - and specialists who can do that are among the scarcest people in industry. So carrier splits, rosters, order quantities, offers and budgets are set by rule of thumb, and nobody learns how far from the best they are.

The Optimization Engine takes the question in plain language and writes the formal model - what to decide, what must hold, what to minimize or maximize - in terms the person asking can check. A solver does the mathematics: exact methods where optimality can be proven, heuristics with a stated quality gap where scale demands them. If the requirements contradict each other, it says so and names them.

The language model is the interface. Solvers are the engine room.

The Optimization Engine turns Prediction Engine output into constrained, actionable decisions. Start with the Analytics Engine, live today and connected read-only to systems already in place.

Request a conversation