The engineers who wrote it are retiring. Some have already died. Yet 250 billion lines of COBOL code continue humming away inside banks, insurance companies, and government agencies—a monument to software that was never meant to last this long.
Now a startup barely out of college thinks it can solve what IBM, Google, and decades of migration vendors could not.
Haladir emerged from Y Combinator's latest Winter batch in early February with Rosetta Docs, a mainframe documentation tool aimed squarely at companies still running systems built in the punch-card era. It's the visible tip of what the San Francisco company describes as a broader AI-powered platform: one that promises not just to document COBOL codebases, but to translate them into Java and C#—correctly.
That last word matters more than you might think.
The Problem Nobody Wants to Inherit
Here's the situation Haladir is betting on: somewhere inside a Fortune 500 company, a 30-year-old system processes millions of transactions daily. Nobody fully understands how it works. The original architects left years ago. The comments in the code are sparse or nonexistent. And every attempt to modernize it has either failed spectacularly or been quietly shelved after someone calculated the risk.
Rosetta Docs tackles the documentation piece first—generating automated explanations of COBOL systems so teams can at least understand what they're dealing with. The company offers a free tier covering up to 10 projects, accessible through a request portal at rosetta.haladir.com. A demo video appeared on LinkedIn showing the tool parsing legacy code, though Haladir hasn't published detailed specifications beyond the portal itself.
Straightforward enough. But documentation is just the wedge.
Behind it sits translation software that, according to Haladir's Y Combinator profile, migrates COBOL and other legacy languages to modern targets while preserving business logic through something the company calls "property-based testing and formal verification." They're using custom models trained with their own approach—RLFR, or Reinforcement Learning from Formally-Defined Rewards, detailed in a recent technical paper.
Which sounds impressive until you remember that code translation tools have existed for 20 years. The real question: does this one actually work?
Formal Methods Meet Silicon Valley Hype
Haladir's technical bet diverges from the typical large-language-model playbook. Their RLFR method fine-tunes open-source models (Qwen2.5 Coder-7B and Qwen3-8B, for those keeping score) using Dafny, a formal specification language, paired with the Boogie verifier and Z3 theorem prover. Train the model on formally verified code, the theory goes, and it learns not just to generate syntax but to preserve correctness.
The company claims improved performance on general coding benchmarks despite training exclusively on Dafny—suggesting the verification discipline transfers to other languages. Whether this translates to provably correct COBOL-to-Java transformations at enterprise scale? That's the part Haladir hasn't demonstrated with public benchmarks or customer case studies.
Still, the theoretical appeal is undeniable. Mainframe migrations don't fail because of syntax errors. They fail when subtle logic bugs—the kind that only surface three months into production—bring down transaction processing on a Tuesday afternoon.
The startup also open-sourced MOBOL, a Python SDK for automating interactions with UniKix COBOL systems on AWS EC2. It's essentially a testing harness, designed to validate that translated code behaves identically to the original through automated CICS transaction testing. The tool requires Python 3.13+ and AWS credentials, targeting the narrow slice of enterprises that have already migrated COBOL workloads to Amazon's Mainframe Modernization service.
Narrow, yes. But perhaps deliberately so.
An Intensely Competitive Landscape
The field Haladir just entered is already crowded with players who've spent years—sometimes decades—building similar tools.
IBM's watsonx Code Assistant for Z became generally available in late 2023, offering discovery, explanation, refactoring, and COBOL-to-Java transformation. The company expanded it with new z17 integration earlier this year. Google announced its own Gemini-powered modernization suite last April. CloudFrame launched "Atlas" in September 2025, with a free tier analyzing 200,000 lines of code.
Then there are the legacy modernization specialists: TSRI with its JANUS Studio (claiming 250+ completed projects), Heirloom Computing, Raincode, Micro Focus. These vendors have been automating COBOL translation since before large language models became a thing.
So what's different now? Haladir's implicit argument seems to be that LLMs plus formal verification create something genuinely superior—not just the same automation with a fresh coat of AI paint.
The market dynamics, at least, support the attempt. A Government Accountability Office report from May 2023 identified 10 critical federal legacy systems costing roughly $337 million annually to operate, with mounting security and staffing risks. Industry research cited by ITPro pegged enterprise losses from legacy technical debt at an average of $370 million per year. IBM's estimate of 250 billion lines of production COBOL hasn't budged.
Even Morgan Stanley recently built an internal AI tool (DevGen.AI) to translate legacy code into readable documentation—proof that major financial institutions see value in AI-assisted understanding, if not yet full automated rewrites.
A Team Straight Out of College
The founding team skews young. CEO Jibran Hutchins comes from Carnegie Mellon. Co-founders include Joseph Tso from Princeton, Preston Schmittou (described as a freshman at UVA Wise—yes, a freshman), and Quan Huynh. According to CB Insights, the company raised roughly $500,000 via convertible note, with seed participation from Y Combinator, Susa Ventures, Joshua Browder (the DoNotPay founder), and SV Angel.
Haladir presented at FinovateEurope 2026's Impact Zone in London on February 10-11, sharing the stage with other financial services startups. Their YC Demo Day is scheduled for March 24, which should provide clearer signals on traction and whether anyone is actually using this beyond the free tier.
Show Me the Proof
Which brings us to the central problem every COBOL modernization vendor faces: credibility.
Translation tools have existed for decades. Some work reasonably well on simple codebases. Most struggle with the gnarly, business-critical systems that companies are actually afraid to touch. The question isn't whether Haladir's technology is clever—it probably is—but whether it works reliably on the kind of code that keeps CIOs awake at night.
Haladir's emphasis on formal verification and property-based testing suggests the founders understand this credibility gap. What they haven't published: customer logos. Case studies. Independent benchmarks demonstrating correctness on real-world codebases at scale. The sorts of things that would let an enterprise architect justify betting a mission-critical migration on a startup founded last year by people who look like they just finished their undergrad theses.
What Haladir does offer is a free documentation tool that enterprises can test without signing anything major—a sensible wedge into accounts that might eventually need full translation. The MOBOL SDK, released as open source, serves double duty: genuinely useful for teams working with AWS's Mainframe Modernization service, and a demonstration that these founders take the technical work seriously.
Whether Rosetta Docs becomes the entry point for larger modernization contracts, or just another piece of an already cluttered toolkit, depends entirely on execution. The technical approach looks interesting on paper, maybe even promising.
Now comes the part where someone has to bet their COBOL migration on it. That's when we'll know if this is different, or just the latest attempt to solve a problem that has outlasted most of the people who created it.
