In this interview, Eqvista speaks with Andrew G. A. Correa, Founder of Monostate, an AI company focused on making it easier for teams to build and deploy custom, high‑performance models for specialized tasks. With a background in artificial intelligence research and hands-on experience turning advanced ideas into practical tooling, Andrew is helping redefine how companies think about model training, fine‑tuning, and inference at scale.
In our conversation, we explore his journey to founding Monostate, his perspective on the future of “stateful” AI systems, and how emerging AI infrastructure connects with modern equity management and startup building—key themes at the heart of Eqvista’s mission to support founders and high-growth companies.
Andrew, you’ve built a fascinating career path, from operations to founding Monostate. Could you take us through that journey? What was the moment that made you realize the industry needed to move away from general-purpose models toward specialized, stateful AI systems?
I started working at 14 in Brazil—office jobs, imports, then eventually leading tech, product, and operations at furniture and tech companies. I scaled two companies to around $100 million in revenue each. But every time I hit this wall where the company grows and suddenly you’re drowning in KPIs, processes, culture stuff… I realized I’m more of a startup person. I love the chaos of growth, not the maintenance phase.
When I left my last company in March, I went straight to the US to understand what was really happening in AI. I did hackathons, met people, connected dots. The realization about specialized models came from two places: first, my research. I published a paper in August on entropy-guided refinement—basically a technique where we capture token-level uncertainty and use it to make smaller models perform nearly as well as reasoning models at a fraction of the cost. We’re seeing small models approach 95% of a reference reasoning model’s quality at roughly one-third the cost.
But the deeper realization is almost philosophical. Look at GPT with 800 million weekly active users—the most successful consumer product in history. That’s also a ticking time bomb. The bias that can be embedded there, the power one company holds over how hundreds of millions of people think… history tells us terrible stories about moments when companies or governments tried to control that many people. This is creeping into people’s lives quietly, with free plans, and people aren’t seeing it. I believe the future should be personal robots—like R2-D2 from Star Wars—not a single global mind that sees everything.
For executives in traditional industries who might not be familiar with Monostate yet, how would you explain what you’re-building in plain language? What’s the core problem you’re solving that keeps your customers up at night?
In simple terms, we help companies create their own AI models instead of renting someone else’s. Think about it—right now, if you want AI in your business, you’re basically paying rent to OpenAI or Google, sending your data to their servers, and getting a one-size-fits-all solution.
The problem that keeps our customers up at night is threefold: cost, because these API calls add up fast at scale; quality, because a general model doesn’t understand your specific domain; and control, because your data and your competitive advantage are flowing through someone else’s infrastructure.
We’re building a platform where anyone—even without engineering experience—can train specialized models for their specific needs. A mom wanting to find the best private schools for her kids. A bank wanting to score credit risk. A hospital wanting to analyze medical imaging. Each gets a model trained for exactly what they need.
Your materials emphasize that ‘specialist beats generalist every time’, with multiple small models orchestrated together instead of one massive model. From an industrial perspective, where do you see the breakeven point—tasks where a specialist model clearly wins versus where a general model is still the rational choice?
The breakeven is really about volume and specificity. If you’re running inference on something like 20 million tokens per month or less, and your use case is generic—summarization, basic Q&A, general writing—stick with a general model. The economics don’t make sense otherwise. But the moment you have volume and domain specificity, the math flips dramatically. We’re consistently seeing 3x cost reduction for inference when we train specialized models for clients, considering server costs to keep the model online. And it’s not just cost—quality improves because the model actually understands your domain deeply. There’s also a speed advantage people don’t talk about enough. With small specialized models,we can test new architectures and iterate incredibly fast. You can’t do that with a 400-billion parameter model. The future isn’t one giant brain—it’s an orchestra of specialists working together.
On the training platform, you promise ‘zero engineering required’ to train and deploy custom models, from risk scoring to translation to medical imaging. Under the hood, what is the minimum ML and data maturity an enterprise actually needs to succeed with Monostate?
Our goal is literally zero experience required. We’re launching our platform in the coming weeks—January 2026—and we’ve built it so that even someone with no technical background can train models.
How? We’re developing techniques for synthetic data generation—data created by capable language models with proper guardrails—plus augmentation of existing data and automatic organization into training datasets. So even if you come to us with messy, incomplete data, we can work with that. That said, for enterprise deployments, you obviously get better results faster if you have clean, structured data. But we’re not going to turn away a mom who wants to train a model to find the best schools for her kids just because she doesn’t have a data science team. That’s the whole point.
How do you envision partnerships with LLM providers, cloud platforms, and developer tools?
We’re currently part of the AWS and NVIDIA startup programs, which gives us access to compute resources. As we scale training volume, we’ll pursue deeper partnerships with GPU cluster providers to bring better costs to our platform—so our customers can train more, for longer. On the tooling side, we have an initial partnership with Weights&Biases for training observability, and we’re pursuing a partnership with Hugging Face. Who knows, maybe even OpenAI in the future if they lean more into open-source models.
The vision is to be infrastructure-agnostic. We want customers to train on whatever cloud makes sense for them, with best-in-class observability, and seamless deployment options.
You claim specialist models trained through Monostate are 3× cheaper than GPT-5 for targeted tasks. What optimizations or model-selection strategies make that possible?
It comes down to three things: smaller models that actually fit the task, efficient serving infrastructure, and smart training techniques.
First, when you train a specialist model, you’re not paying for 400 billion parameters when you only need 7 billion. Most business tasks don’t need frontier model capabilities—they need deep domain expertise.
Second, we use serverless inference providers and optimize serving based on each client’s volume patterns. You’re not paying for idle GPUs.
Third, our entropy-guided refinement technique. We published research showing that a small model with our uncertainty-aware loop approaches 95% of a reasoning model’s quality at approximately one-third the cost. We achieve selective refinement on about 31% of responses while improving accuracy by 16 percentage points over single-pass inference. It’s the middle ground between cheap-but-dumb and expensive-but-smart.
Industrial buyers care less about AI demos and more about uptime, audit trails, and liability. How do you design workflows and guardrails—validators, ensembles, or human-in-the-loop—to reduce hallucinations and meet compliance expectations in regulated sectors?
This is core to our research. Our published work on entropy-guided refinement is essentially a hallucination detection system. We extract token-level probabilities, compute Shannon entropy on alternatives, and when the model is uncertain, we pass an uncertainty report back to it for corrective edits. The model essentially tells you”I wasn’t confident here” before it becomes a problem.
For regulated sectors—hospitals, financial institutions—we’re building an on-premises version of our platform. These clients can’t send data outside their protected environments, and we respect that. They can train all types of models—not just LLMs, but regression, tabular, video, image, audio—within their secure, regulated environments.
The key insight is that compliance and capability aren’t trade-offs. If you build uncertainty awareness into the model itself, you get better outputs and better audit trails.
If you were advising a traditional industrial company about its first serious AI initiative in 2025, what would you tell them to do in the next 90 days, and what common pitfalls would you warn them to avoid?
Talk to your people. Seriously. This is the best thing most companies will never do, and it’s the biggest mistake I see.Here’s what usually happens: company decides to “do AI,” they grab some customer support data, train a model, deploy it, and then discover it doesn’t actually work for the job. Then they iterate, guess again, retrain, redeploy… it’s expensive trial and error.
Flip it around. If you want to automate a job, why would you try to guess how people do it when you can just ask? Go talk to the humans who do that job well. Understand their process, their edge cases, the weird situations that only show up on Fridays, the shortcuts they’ve developed over years. That’s your gold.
Once you understand how the best humans actually do the work, you can build the right datasets and the right benchmarks. You’re not guessing anymore—you’re encoding real expertise.
So my advice for the first 90 days: don’t touch a single model. Spend that time with your people. Interview them, shadow them, document what good looks like. Then build your AI to match that standard. The companies that skip this step waste months and millions learning what a few conversations would have told them upfront.
Finally, what upcoming capabilities or product directions are you most excited about that you can share publicly? And for companies that want to experiment with digital collaborators, what’s the easiest way to get started working with you?
We’re launching our “vibe training” platform in the coming weeks—within January 2026. Think of it like Lovable, but for training AI models instead of building apps. Anyone will be able to train small and medium models without writing code. After launch, we’re adding more model types: audio TTS models, experimental architectures like text convolution, and an advanced module where our proprietary Trainer models work alongside advanced users to develop custom architectures and research.
Long-term, I’m most excited about our work on stateful architectures—hybrid systems that update parameters at inference time. Imagine AI that actually remembers conversations naturally, like humans do, with short-term and long-term memory, instead of the current hack of appending previous messages to every prompt.
To get started, reach out directly. We work with enterprises on custom training projects right now, and soon anyone will be able to jump on the platform and start training. We’re building the future where everyone can create their own AI—not rent someone else’s.
