The LinkedIn post was almost aggressively understated. Jibran Hutchins, co-founder of a startup called Haladir, announced in late February that his team had launched something called Rosetta Docs—a documentation tool built specifically for COBOL. There was a demo link. A note that translation features were "coming soon." That was it.
For a four-person outfit fresh out of Y Combinator, it read less like a product launch and more like someone testing the waters. Which makes sense, given what they're up against: IBM, Amazon Web Services, and an entire ecosystem of established modernization vendors who've been circling the mainframe market for years.
But here's the thing about that market. It's enormous, lucrative, and stuck in a time warp that's starting to show cracks.
Mainframes still handle almost 70% of the world's production IT workloads, according to IBM's own research. Forty-five of the top 50 banks depend on them. So do 67 of the Fortune 100. And the overwhelming majority of that code? COBOL. Written decades ago. Often by developers who have since retired, taken their institutional knowledge with them, or both. The documentation, when it exists at all, tends to be incomplete, outdated, or maddeningly vague.
Which brings us back to Haladir, and why a tiny team thinks they've found an opening.
The Problem That Won't Go Away
Enterprises looking to modernize mainframe systems keep running into the same wall: you can't migrate what you don't understand. Legacy COBOL codebases can span millions of lines, with business logic embedded in ways that aren't immediately obvious to engineers trained on Java or Python. Reverse-engineering that logic by hand is slow. Expensive. Sometimes impossible.
Rosetta Docs is Haladir's attempt to automate that initial step. According to the company's Y Combinator profile, the tool uses AI to generate comprehensive documentation from COBOL codebases. There's a free tier for up to 10 projects. Enterprise inquiries go to an email address. Pricing beyond that? Not public yet.
The documentation product, though, is just the first move. Haladir's YC profile also describes a translation tool in the works—one that's designed to migrate COBOL and other legacy languages to Java and C# .NET. The company says it uses "property-based testing, formal verification, and custom models" to ensure accuracy. That piece isn't live yet. Hutchins noted in his launch post that updates are coming, but offered no timeline.
A Market Already Packed With Players
Haladir is hardly breaking into virgin territory. IBM's watsonx Code Assistant for Z—which Big Blue has, without irony, called a "Rosetta Stone for mainframes"—already offers code understanding, refactoring, and COBOL-to-Java transformation. An IBM Research blog post from late June 2025 cited a client case study with a 60% productivity increase, though the details were sparse.
AWS is pushing aggressively too. In a March interview with ITPro, AWS Vice President Asa Kalavade described the company's Transform for Mainframe platform, which uses what he called "agentic AI" to translate COBOL, JCL, CICS, and other legacy languages into Java. But Kalavade was refreshingly candid about the limits: "You need those experts to even define what these transformations are." Translation: the AI accelerates the work, but it doesn't replace the people who actually understand the systems.
Then there's CloudFrame, which launched its Atlas platform in September 2025. The company pairs system intelligence with what it describes as "deterministic, AI-augmented" COBOL-to-Java conversion. Free tier: 200,000 lines of code per project. Rocket Software and Heirloom Computing are in the mix as well, alongside consulting-heavy services from firms like Kyndryl and Deloitte.
In other words: crowded.
The Formal Methods Bet
What separates Haladir—at least on paper—is its focus on formal verification. The company has published research on reinforcement learning from formally-defined rewards for code generation, using tools like Dafny and Z3 to verify program correctness. Another project, ConstraintBench, benchmarks how well large language models handle constrained optimization problems. (Answer: not well. According to a company LinkedIn post, no model scored above 31% on tasks that were both feasible and optimal.)
The formal verification angle isn't just academic posturing. Mainframe code often runs critical infrastructure—payment systems, regulatory compliance checks, the kind of stuff where a 98% success rate isn't remotely good enough. If the 2% involves miscalculating a mortgage payment or mangling a healthcare claim, you've got a problem. Whether Haladir's approach can deliver that level of assurance in production environments remains an open question. The company hasn't named any public customers yet. Or partners, for that matter.
What Comes Next
The founding team has the pedigree: Carnegie Mellon, Princeton, UVA Wise. They've got backing from Y Combinator, Susa Ventures, SV Angel, and Joshua Browder, who posted on social media that he was "proud to be the first investor." A CBInsights tracker from January listed a $500,000 convertible note, though a subsequent LinkedIn update mentioned additional investors—suggesting the total raised may be higher, though the company hasn't confirmed that.
For now, Rosetta Docs is the only product you can actually use. The translation tool is still in development, and Haladir's website leans heavily on research rather than go-to-market messaging. That's fairly typical for an early-stage YC company. But it also means the hard part—proving the technology works in real-world, high-stakes environments—is only just beginning.
Enterprises move slowly, especially when legacy systems are on the line. Haladir's strategy seems to be to use documentation as a wedge, a way to get inside organizations before pitching the riskier proposition of actually translating their code. It's a sensible approach, maybe even the only practical one for a four-person team going up against IBM and AWS.
Whether it's enough is another question entirely.
