Pramaana Labs Raises $27 Million to Build a Formal Verification Layer for AI in Tax, Healthcare and Compliance
In April 2026, a researcher cataloguing AI errors in legal proceedings counted over 1,353 documented hallucination cases in courtrooms worldwide. On a single day, ten new cases emerged from ten different courts. In the first quarter of 2026 alone, US courts issued more than $145,000 in sanctions against attorneys who submitted AI‑fabricated citations. Sullivan and Cromwell filed a bankruptcy motion containing roughly 28 erroneous citations. A Department of Justice attorney faced sanctions after a government brief was found to contain fabricated quotes and nonexistent regulatory language. In parallel, a clinical review found that over 90% of clinicians surveyed had encountered hallucinations from medical AI, with 85% considering them capable of causing patient harm.
These are not fringe incidents. They are the accumulating evidence base for a problem that Pramaana Labs was founded to solve architecturally. On June 17, 2026, the company announced a $27 million Seed round led by Khosla Ventures, with participation from Accel, BoldCap, Nexus Venture Partners, Premji Invest, and Unbound. The capital will be deployed toward training formalisation and prover models, expanding the AI research team, and onboarding domain experts across regulated sectors including taxation, healthcare diagnostics, cybersecurity, and financial compliance.
The Core Thesis: Proof, Not Probability
Most AI systems in commercial deployment today operate probabilistically. A large language model does not know whether its answer is correct it generates the response that is statistically most likely given its training data and the query it received. For a wide range of applications, that is acceptable. For the domains Pramaana is targeting tax, law, clinical medicine, financial compliance it is not.
Pramaana's architecture addresses this limitation at the infrastructure level rather than at the prompt or guardrail level. The system works in three stages. First, rules and regulations tax codes, clinical protocols, legal statutes, financial compliance frameworks are translated into a formal language that machines can process with logical rigor. Second, when a user submits a query, the system translates that query into a formal statement and runs it through a proof engine. Third, the system either returns a machine‑checkable proof supporting its answer, or it identifies the specific rule that prevents a valid conclusion from being drawn. If no proof can be established, the system declines to respond.
The technical foundation draws on LEAN, an open‑source programming language originally developed for verifying mathematical proofs. Pramaana applies this framework to domains where knowledge is rule‑governed in a similarly structured way. Ranjan Rajagopalan, co‑founder and CEO, draws the analogy explicitly: tax law, like mathematics, has a codified rule set. Once that rule set is formalised, reasoning on top of it becomes deterministic rather than probabilistic. The output is no longer a confident‑sounding guess it is a derivation with a traceable chain of logic that can be audited against the source rules.
The company points to France's CATALA project as a precedent for the underlying approach. CATALA has formalised significant portions of the French tax and benefit system into executable code, demonstrating that statutory language can be rendered machine‑processable without losing legal meaning. Pramaana's ambition is to build that infrastructure across multiple high‑stakes domains simultaneously, with domain experts embedded into each vertical to ensure the formalisation is legally and technically accurate.
The Founding Team and Domain Coverage
Pramaana Labs was founded in September 2025 by Ranjan Rajagopalan, Krishnan Raghavan, and Sanjay Ganapathy Subramaniam, all IIT Madras alumni. The company is registered in Bengaluru and maintains a presence in San Francisco. It held its inaugural Verification Summit on June 10, 2026, in San Francisco, headlined by Vinod Khosla himself an unusual level of public involvement from a lead investor at the seed stage, and a signal of the conviction behind the investment.
For the tax law vertical, Pramaana is working with Danny Werfel, the former IRS Commissioner. Professors from IIT Delhi, IIT Madras, and UC Berkeley oversee the cybersecurity and drug discovery systems. The roster of domain experts is not ornamental formalising a tax code or clinical protocol requires both deep subject‑matter knowledge and the technical capacity to translate that knowledge into machine‑verifiable representations. The two competences rarely coexist in the same person, which is why the company is building a structure that pairs AI researchers with domain specialists across each vertical.
The Hallucination Problem at Scale
The problem Pramaana is targeting has quantifiable scope. AI hallucination rates on legal queries have been measured at 88% in some evaluations. In legal proceedings, that translates directly into professional sanctions, wasted court time, and at the downstream end miscarriages of justice. In healthcare, the consequences are clinical. In tax and financial compliance, errors carry regulatory penalties, reputational damage, and fiduciary liability.
The broader regulatory context sharpens the commercial urgency. The EU AI Act established a compliance framework for high‑risk AI systems operating in law, healthcare, and financial services, originally targeting August 2026. A provisional Omnibus amendment agreed in May 2026 by the European Parliament and Council deferred most of those requirements to December 2027. The deferral does not alter the direction of travel. Enterprises deploying AI in regulated domains will face mandatory accuracy, robustness, and human oversight requirements that probabilistic systems cannot meet through prompting or post‑hoc guardrails. Formal verification, in which correctness is demonstrated rather than estimated, is architecturally positioned to meet those requirements.
What the Capital Signals
A $27 million seed round is large in absolute terms but modest in the context of AI infrastructure plays. Frontier model companies have raised hundreds of millions at formation for compute‑intensive training. Pramaana's raise reflects its infrastructure‑layer positioning the company is not training a foundation model but building a verification layer that sits above existing LLMs. The approach is closer in spirit to zero‑knowledge proof systems in cryptography, where trust is mathematical rather than institutional, than to conventional AI model development.
Sathya Narayanan, general partner at BoldCap, described auto‑formalisation infrastructure as a potential critical building block for future AI systems, framing the investment as a bet on the foundational layer of trustworthy AI rather than any single application. Vinod Khosla characterised auto‑formalisation as a missing capability a gap in what current AI systems can do reliably rather than an incremental improvement on existing approaches.
The name the founders chose carries its own signal. Pramaana is a Sanskrit term meaning a valid means of knowledge. In Indian philosophical tradition, it refers specifically to sources of knowledge that are reliable enough to ground justified belief. The company's founders have named their venture after the concept of epistemic validity itself a statement of intent, and also a succinct description of what enterprise AI currently lacks and what regulated industries cannot function without.
