Imagine preparing a legal argument under pressure, with a tight deadline, and high stakes, while you rely on an AI tool to surface the right precedent. It gives you a confident answer, complete with citations. Everything looks right. Except one of those cases doesn’t exist.
This is the problem Rilton Franzone is trying to solve.
Franzone began teaching himself computer science at 15 through online courses from institutions like Harvard, MIT, and HKUST. His trajectory is defined by speed and focus. Today, as a software engineer at Midpage.ai, he is working on a harder problem than building AI that works: building AI that can be trusted when mistakes carry real consequences.
A Builder Shaped by Real Systems
Franzone’s path into technology was not defined by formal structure, but by early exposure to real-world problems. By the time he was 17, he worked full-time at a startup, moving across backend, frontend, mobile, and data systems.
That breadth wasn’t accidental. It forced him to confront how systems actually behave outside of controlled environments.
Across roles in fintech, academic tooling, and SaaS, he worked on infrastructure processing millions of loan applications, built crawlers indexing hundreds of thousands of machine learning implementations, and contributed to platforms serving large user bases across industries.
What he saw, repeatedly, was this: systems rarely fail loudly. They fail quietly.
An edge case slips through. A data pipeline introduces a subtle inconsistency. An output looks correct, but isn’t. These are the kinds of failures that don’t crash systems, but slowly erode trust.
That pattern would later shape how he approaches artificial intelligence.
Building Trust at Midpage.ai
When Franzone joined Midpage.ai as its third engineer, before the company’s seed round, it was generating roughly $90K in annual recurring revenue.
But the more important signal wasn’t growth, it was adoption under scrutiny.
Today, more than 300 law firms across the United States use the legal research AI agent he built every week, often for hours at a time. The system has also been evaluated in VLAIR’s benchmark from vals.ai, ranking among the top three globally and outperforming both general-purpose AI systems and human lawyer baselines across multiple metrics.
In legal workflows, performance alone is not enough. A tool is only useful if a professional can rely on it without second-guessing every output.
That is a much higher bar.
Franzone has also led integrations with major technology partners, including Perplexity, Litera, Anthropic, and OpenAI, expanding the platform’s reach while contributing directly to its revenue growth.
Why Reliability in Legal AI Is Different
In many domains, an AI error is inconvenient. In law, it is consequential.
A fabricated citation, a mischaracterized ruling, or an invented piece of supporting evidence introduces risk into decisions that carry legal and financial weight.
“My main concern is making sure these tools do not hallucinate,” Franzone explains.
In practice, that means building systems where every claim can be traced back to a real source where outputs are not just plausible, but verifiable.
This changes how success is defined. Accuracy is not enough. Responses must be grounded, explainable, and interrogable by the user.
Measuring What Actually Matters
To tackle this problem more directly, Franzone has been developing benchmark.midpage.ai, an evaluation framework designed to test how AI systems perform on the kinds of tasks lawyers actually face.
Traditional benchmarks tend to isolate narrow capabilities: retrieving a fact, summarizing a document, answering a constrained question.
But legal work doesn’t happen in isolated steps.
It unfolds over time. It involves ambiguous inputs, multiple sources, and extended reasoning that can take days or even weeks to complete manually.
The benchmark reflects that reality. Built in collaboration with practicing lawyers, it evaluates AI systems on complex, multi-step tasks and introduces a structured judging system capable of assessing not just correctness, but relevance and reasoning quality at scale.
The goal is to make limitations visible. In a high-stakes environment you need to know what your systems can’t do.
Operating at the Intersection of Engineering and Product
Franzone’s technical stack spans modern TypeScript, Node.js, React, Next.js, Python, and AI engineering. His strength isn’t in a single tool, but in how he deals with uncertainty.
Working in small, fast-moving teams means he had to be able to move between disciplines, design infrastructure, build product features, debug pipelines, and translate user needs into technical decisions.
In these environments, there is rarely a clear roadmap. The challenge is not just building systems, but deciding what to build, and why.
“I see myself essentially as a problem-solver,” he says, a mindset that has allowed him to adapt quickly as technologies and demands evolve.
Alongside his practical work, Franzone maintains a strong interest in the theoretical foundations of computer science, aiming to bridge hands-on system building with a deeper understanding of the principles behind them.
The Direction of High-Stakes AI
Franzone’s long-term focus is consistent: building AI systems that can operate reliably in professional environments where correctness is non-negotiable.
Legal technology is one of the most demanding testing grounds for this. The field is defined by precision, precedent, and accountability. Any system that participates in it must consistently meet those standards.
As AI becomes more deeply embedded in professional workflows, the question is shifting.
It is no longer just: Can this system generate an answer?
It is: Can you trust it when the stakes are real?
For Franzone, that is the problem worth solving. Follow him on LinkedIn and GitHub.
(Source: GitHub)
