Release v0.1

We are excited to announce the Beta Release of Harumi Platform, now available at platform.harumi.io!

Bringing Optimization and Machine Learning Together

Operations Research (OR) and Machine Learning (ML) are powerful tools used across industries to improve efficiency, reduce costs, and foster innovation. By leveraging data and mathematical models, businesses can optimize operations, enhance decision-making, and drive revenue.

While optimization is everywhere, from solving routing problems to managing inventories, the bottleneck often lies in modeling the problem, not the solver itself. Most businesses require specialized teams to model their optimization problems, and for smaller-scale issues, the cost of such teams may outweigh the benefits. This creates a barrier to widespread optimization, as the key challenge is not in solving the problem but in defining and modeling it correctly.

Making Optimization More Accessible with AI

To democratize optimization, Harumi leverages Large Language Models (LLMs) to automate the modeling process. Our approach includes:

1. Understanding Business Logic in Natural Language: Translating business problems into structured mathematical terms.

2. Mathematical Reasoning: Converting business concepts into linear constraints.

3. Implementing Solver Code: Writing the code for solvers like Gurobi, which optimizes based on the provided constraints.

By bridging the gap between business logic and mathematical modeling, LLMs significantly reduce the need for specialized modeling teams, making optimization more widely accessible.

Why Not Just Use LLMs for Problem-Solving?

A simple approach is to ask a general-purpose LLM (like ChatGPT) to solve an optimization problem. However, this method has key limitations:

• Unreliable Outputs: LLMs may produce solutions with token prediction errors.

• Ambiguous Representations: Complex problems often lack precise descriptions, making them hard to interpret.

• Long Problem Descriptions: Industries like power systems involve extensive datasets that general LLMs struggle with.

• Large Data Matrices: Optimization problems often involve thousands or millions of variables and constraints.

• Hallucinations: LLMs can generate plausible-looking but incorrect results.

By combining LLMs with traditional solvers, Harumi ensures reliable, structured, and precise optimization solutions.

Harumi Framework Architecture

We have developed a cutting-edge AI agent with 10% higher accuracy in solving mathematical and quantitative problems compared to other LLMs, designed to assist with machine learning, operations research, and data analysis. Our AI agent supports optimization modeling making complex decision-making more intuitive. The agent architecture was develop to support:

1. Natural Problem Description: Users describe their optimization problems in plain language.

2. Pre-Processing: Identify parameters, objectives, and constraints.

3. Iterative Modeling: LLMs assist in structuring optimization models with built-in validation checks.

4. Formulator Agent: Converts problem descriptions into a formal mathematical model (e.g., LaTeX).

5. Programmer Agent: Translates this model into solver-ready code (e.g., Gurobi, CPLEX).

6. Evaluator Agent: Detects errors and ensures model accuracy.

Key Features of Harumi Beta

• Create and share optimization projects with complex mathematical models.

• Smart text editor tailored for documenting operations research projects.

• Automatic code generation, integrating with solvers like Gurobi and CPLEX.

• Conversational AI assistant to help start new optimization projects in logistics, infrastructure, and more.

• Enterprise-grade data security, running LLMs on a private cloud without sharing data externally.

The Harumi Beta is now live! Sign up today at platform.harumi.io and explore how AI can revolutionize your optimization and decision-making processes.

Coming Soon: Exciting new features are expected for next releases, stay tuned!

• Execute Python code in the browser with WebAssembly for quick iterations.

• Server-side code execution to support more complex processing.

• Backend integration with various solvers for code execution.

• Develop a visualization and interaction app with graphical components.

• Integrate with external data sources (datatable, Excel, CSV) for seamless code execution.

• Provide a simulation environment for execution results.