Reliability has been in the spotlight as companies struggle to turn AI pilot programs into functional parts of their businesses. A new startup hopes to solve this problem by leveraging a mathematical formalization tool that combines one of computer science’s most reliable systems with one of its most confusing.
Pramaana Labs on Wednesday announced $27 million in seed funding led by Khosla Ventures, with participation from Accel, Boldcap, Nexus Venture Partners, Premji Invest, and Unbound.
Pramaana will focus on highly sensitive areas such as law, drug development, and tax preparation, where errors are costly and reliability is paramount. Deploying AI in these systems will require stronger protection against hallucinations and errors than we currently have. But in the view of Ranjan Rajagopalan, co-founder and CEO of Pramaana, they are also uniquely suited to formalization.
“It’s like math in that there are a lot of rules you have to follow,” Rajagopalan told TechCrunch, explaining the tax code rules. “Once you have a coded version, your inferences from it start to become deterministic.”
Pramaana’s system still runs on traditional LLM, giving it the flexibility to answer natural language questions and solve complex problems that traditional computers cannot handle. But on top of that LLM, there is a crucial layer that confirms the work of the LLM.
The combination of an LLM engine and deterministic verification is a widely used setup. Pramaana’s unique approach is the use of formal verification tools. That is, leveraging the open source LEAN programming language used to verify mathematical proofs. Much of this work has real-world precedent. Rajagopalan points to France’s CATALA project, which formalizes much of France’s tax and benefits system into executable code.
For each use case, Pramaana builds its own LEAN-style formal verification system overseen by domain experts. For tax law, the company is working with former IRS Commissioner Danny Werfel, and the professor at IIT Delhi, IIT Madras and UC Berkeley oversees cybersecurity and drug discovery systems.
“The world’s most difficult problems are not unsolvable. They are unstructured,” says Rajagopalan. “There are rules in every area where if something goes wrong, someone could lose their health, money or freedom.”
Now all we have to do is codify these rules.
If you purchase through links in our articles, we may receive a small commission. This does not affect our editorial independence.
