Pramaana Labs secures $27M to formalize AI reasoning in high-stakes domains

Pramaana Labs, a startup focused on making large language models reliable enough for regulated industries, announced a $27 million seed round on June 17. According to TechCrunch, the round was led by Khosla Ventures and included participation from Accel, Boldcap, Nexus Venture Partners, Premji Invest, and Unbound.

The core insight animating Pramaana’s approach: rule-heavy domains like tax policy, drug approval workflows, and contract law are mathematically formalizeable in ways that generic language tasks are not. By layering formal-verification logic atop a conventional LLM, the startup aims to eliminate hallucinations in contexts where errors cascade into compliance failures or medical harm.

Formal verification as an LLM constraint layer

Pramaana embeds mathematical reasoning into its AI stack by leveraging LEAN, an open-source programming language originally designed to verify mathematical proofs. Rather than replacing the LLM, the company wraps it: the model generates candidate answers to natural-language queries, and a deterministic reasoning engine—built using LEAN-style formal rules—validates that each answer complies with the applicable regulatory framework.

According to TechCrunch, Pramaana CEO Ranjan Rajagopalan compared rule-governed domains to mathematics itself. “Once you have a codified version of [the rules], the reasoning on top of it starts becoming deterministic,” Rajagopalan told the publication. This determinism is the opposite of LLM behavior, where semantic plausibility often masquerades as accuracy.

Domain-specific implementations with expert oversight

Pramaana is not building a one-size-fits-all verification engine. Instead, for each vertical, the startup will construct its own formal ruleset, overseen by subject-matter authorities. The company has enlisted former IRS Commissioner Danny Werfel to guide its tax-preparation work, while professors from IIT Delhi, IIT Madras, and UC Berkeley are supervising drug-discovery and cybersecurity applications.

This approach mirrors France’s CATALA project, which Rajagopalan cited to TechCrunch as precedent—a real-world effort to encode French tax and welfare rules into executable, verifiable code. The playbook is proven; the novelty lies in integrating it with LLM-based natural-language interfaces.

Why this matters

Enterprises have struggled to move AI pilots into production precisely because hallucination risk is unacceptable in high-consequence domains. Regulatory bodies are beginning to demand explainability and verifiability as conditions for approval, especially in pharmaceuticals and financial services. By reducing AI outputs to checkable logical propositions rather than probabilistic text generation, Pramaana addresses both the technical and compliance sides of that friction.

The startup’s success will depend on whether formal-verification overhead—the engineering effort to codify each domain’s rules—scales to new verticals faster than manual review. If it does, the model could unlock billions in cautious enterprise AI spending currently sitting on the sideline.