I keep hearing the same confession from executives who have already “done AI”: the pilot worked, the demo dazzled, the board nodded—and then reality arrived. The model that sounded so confident in a sandbox suddenly had to live inside a business where mistakes cost money, lawsuits, reputations, and sometimes people’s lives. That’s when the conversation stops being about innovation and starts being about a less glamorous word: reliability.
And here’s the uncomfortable question: if your AI can’t be trusted in the moments that matter, what exactly did you build—an assistant or a liability?
When pilots don’t become products
The gap between “cool prototype” and “production system” is widening, not shrinking. It’s not because companies lack talent or compute. It’s because large language models, by design, can be brilliant and wrong in the same sentence. They can improvise. They can hallucinate. They can sound authoritative while being completely untethered from the rules your industry is built on.
That’s why I’m paying attention to startups trying to weld structure onto that chaos instead of just turning up the temperature knobs and hoping for the best. One of them is Pramaana Labs, which recently announced a $27 million seed round led by Khosla Ventures, with participation from Accel, BoldCap, Nexus Venture Partners, Premji Invest, and Unbound.
The pitch is straightforward in spirit and ambitious in execution: combine the flexibility of an LLM with the determinism of formal verification—a corner of computer science that cares less about eloquence and more about proof.
Why “high-stakes” industries are the real test
Pramaana is aiming at domains where the cost of being wrong isn’t merely a bad customer experience. Think law, drug discovery, and tax preparation—areas where a single error can trigger audits, derail clinical decisions, or change the outcome of someone’s freedom or finances.
These fields also share something else that’s easy to forget when we talk about AI like it’s magic: they’re governed by dense rule systems. If you’ve ever tried to interpret a tax code update or track the exceptions to the exceptions in compliance, you already know. It’s maddening, but it’s not random.
Ranjan Rajagopalan, Pramaana’s co-founder and CEO, frames it in a way that lands: tax rules are “like math” in the sense that there are many constraints you must follow. And once those rules are encoded, reasoning over them starts to become deterministic. That’s the heart of the bet: the scariest AI use cases might also be the most “formalizable.”
A deterministic layer on top of a language model
To be clear, Pramaana isn’t rejecting LLMs. The company still relies on a conventional language model because that’s the piece that can interpret natural language questions and navigate complexity in a human-facing way. If you want a system that can talk to a lawyer, a researcher, or a taxpayer, you probably need the linguistic interface.
But Pramaana’s approach adds a second layer above the LLM—one that’s meant to ensure the model’s work “checks out.” In other words, the LLM can propose an answer or a chain of reasoning, but the deterministic layer is there to verify whether the steps actually follow from the encoded rules.
This general architecture—LLM plus verification—has become a popular direction across the industry. What makes Pramaana distinct is the toolset it’s borrowing: formal verification methods that are typically associated with proving the correctness of software or mathematical statements, not wrangling chatbots into behaving.
LEAN, proofs, and the attempt to tame chaos
Pramaana is leaning on the open-source programming language LEAN, which is widely used for checking mathematical proofs. If you’re not steeped in this world, the attraction is simple: LEAN doesn’t care how persuasive you sound. It cares whether the proof is valid.
There’s also precedent for bringing this kind of rigor into bureaucratic reality. Rajagopalan points to the French project CATALA, which has worked to encode large parts of France’s tax and benefits system into executable code. That matters because it suggests the vision isn’t purely theoretical. Governments and researchers have already tried turning sprawling legal text into something a computer can run and verify.
Pramaana’s plan is to build a LEAN-like formal verification system tailored to each use case, with domain experts supervising the translation of real-world rules into formal logic. And yes, that translation is the hard part. It’s one thing to say “encode the rules.” It’s another to do it without smuggling in ambiguity—or accidentally dropping the exceptions that make the rule meaningful.
Experts in the loop, not as a slogan
In tax, Pramaana is working with former IRS commissioner Danny Werfel. In cyber and drug discovery, the company says professors from IIT Delhi, IIT Madras, and UC Berkeley are overseeing the work. That detail matters because it hints at a more serious posture than the usual “we’ll fine-tune and see.” If you’re going to formalize a field, you need people who know where the bodies are buried—the edge cases, the contradictions, the institutional interpretations that aren’t obvious from the text.
I’m also struck by how this flips a common assumption. We often treat regulated, rules-heavy sectors as “hard for AI” because they’re complex. But Pramaana’s worldview is that these sectors are hard precisely because we’ve left so much of their logic implicit—stuck in PDFs, precedent, and professional judgment rather than formal systems.
“Not unsolvable—unformalized”
Rajagopalan’s line is the kind that sounds like a slogan until you sit with it: “The hardest problems in the world aren’t unsolved. They’re unformalized.” He argues that any domain where being wrong can cost someone health, money, or freedom has rules—lots of them. The opportunity is to capture those rules in a form that machines can reason over reliably.
Whether you buy that depends on how you view the messy parts of human systems. Are they messy because nobody bothered to formalize them, or because they can’t be fully formalized without losing something essential? My guess is: both are true, depending on the slice of the problem.
Still, the direction feels right. If we want AI to stop being a high-gloss improviser and start being a dependable coworker, we’ll need more than “better prompts.” We’ll need guardrails that can actually prove something.
What I’ll be watching next
If Pramaana succeeds, it could change what “AI adoption” looks like in high-stakes workflows. Not just “the model answered,” but “the model answered and the system can verify why.” And that’s a different kind of trust—less emotional, more engineered.
- How much of each domain can truly be formalized without collapsing under edge cases and shifting interpretations.
- How the system handles ambiguity, especially where the “right” answer depends on context, precedent, or probabilistic evidence.
- Whether verification slows things down enough to frustrate real users—or whether it becomes the feature people are willing to wait for.
In the end, the promise is almost old-fashioned: write the rules down, encode them carefully, and make the machine show its work. Maybe that’s what the AI era needs most—not more imagination, but more discipline.
Comments