APPLIED DECISION SYSTEMS LAB
We build tools that combine mathematical modeling, machine learning, NLP, and agentic workflows to make complex decisions more understandable and more actionable.
Mathematics is a precise language,
but it only becomes real
when it enters use.
Like in The Velveteen Rabbit, something becomes real when it enters a relationship with the world. In technology, a model becomes real when it informs a workflow, supports a decision, or helps someone act with better judgment.
The Velveteen Project is a founder-led applied decision systems lab. It exists to turn rigorous ideas into usable tools.
We work at the intersection of mathematical modeling, machine learning, natural language processing, and agentic systems to build tools for problems shaped by uncertainty, fragmented information, and operational friction.
We are not interested in AI as spectacle. We are interested in AI as infrastructure: explainable, deployable, and useful.
This project is about moving from theory to operation without losing rigor in the process.
Selected Projects
ECOAGENT
AI-powered stochastic climate risk platform for real-time landslide risk prediction.
- PROBLEM:
- Static risk maps fail when conditions change rapidly.
- CORE:
- CIR stochastic model for soil moisture + real-time weather reasoning.
- INTEGRITY:
- Production architecture separating deterministic models from LLM orchestration. Tool-first anti-hallucination approach.
STOCHASTOGREEN
Climate transition risk simulator for financial portfolios.
- PROBLEM:
- Existing transition models are opaque, hard to update, and lack transparent reasoning.
- CORE:
- Merton jump-diffusion model mapping climate beta via deterministic simulation.
- INTEGRITY:
- Math engine and AI layer intentionally decoupled. The LLM interprets the trajectory; it never computes the math.
From Analytical Core to Usable System
A practical process for turning complex ideas into tools that can actually support real decisions.
FRAME THE DECISION
We identify the decision that matters, the constraints around it, the available data, and the failure modes. The goal is not to automate everything, but to clarify what must be modeled and what must remain visible.
BUILD THE ANALYTICAL CORE
We design the core of the system around the problem itself: a mathematical model, an ML pipeline, or a hybrid architecture when the problem calls for it.
WRAP IT IN WORKFLOW
We add interfaces, retrieval, orchestration, or agentic components only where they improve usability, reasoning, or execution.
TEST FOR USEFULNESS
We evaluate not only performance, but interpretability, robustness, and whether the system actually helps people decide or act better.
Behind the Project
A founder-led lab built around modeling, machine learning, NLP, and decision systems.
Carlos Manuel Orrego
Applied mathematician, computer scientist, educator, and AI builder. I created The Velveteen Project as a public lab for building systems that connect rigorous thinking with real-world use.
CURRENT FOCUS
Building better
decision tools
This project grows through carefully chosen prototypes, technical writing, and real systems built with rigor. The goal is not to look larger than it is, but to become more real with every iteration.
START A CONVERSATIONOperational Philosophy
How we approach the transition from theory to deployed system.
Analytical Core Before Interface
The math comes first. We do not use LLMs to hallucinate numbers. We use deterministic engines for truth. We use agents for reasoning and routing. Interfaces are built to expose the model's logic, not obscure it.
Tools Follow the Problem
We adapt the architecture to the failure modes of the decision. Models remain interpretable. Orchestration layers exist to improve usefulness and constraint handling, not to add theatrical complexity.
Deployment Matters
Rigorous ideas die in Jupyter notebooks. Real tools run defensively in production. We architect for testability, telemetry, and robustness from day one.
Have a decision problem worth
modeling?
If you are working on a complex workflow, a scientific tool, or an applied AI system that needs more than a chatbot layer, I'd be glad to hear about it.
What I Build
- → Applied decision systems
- → Mathematical modeling and simulation
- → ML and NLP pipelines
- → Agentic workflows for complex processes
- → Retrieval and structured reasoning systems
- → Scientific and analytical interfaces
- → Research-driven prototypes with real deployment paths