I co-founded jhana and run its operations. jhana runs in 9+ courts and 5+ government offices, and 400 judges and 15,000+ lawyers use it. We help courts leapfrog: for many of them, the last technology they bought was a LAN cable. Our interventions cut time to numbering and first listing by up to 30×, and rote work by 83%. This looks like an undertrial prisoner can be heard the day their matter is filed. India's backlog is not a capacity problem.

We also run similar AI transformation for law firms and businesses. Through my work, I try to understand the economics of service-based industries in the AI era, and where human judgment and retention become critical.

My dream is to change how a citizen experiences the State. A Zomato-pilled generation still gets legal/ state-services that feel stuck in the '80s.

Previously

• Worked in the chambers of Justice Aravind Kumar, Supreme Court of India. Learned how much of a judgment depends on what actually reaches the judge's desk.
• Started LiNC at Jindal to build talent density around legal AI, for students who wanted to build the law, not just study it.
• Founded Rubd, a vegan cosmetics brand, at 19. Took it to $100K in revenue during law school, then shut it down to finish my degree.
• Have been thinking about techno-fascism since 2019, back when it sounded paranoid.

On my mind

• What does an accepted answer cost? Tokens per answer is the wrong unit; review, correction and re-runs are most of the bill. (Kapoor et al., 2024)
• Can RL with verifiable rewards work where the right answer is a judgment call? Law has very few unit tests. (Lambert et al., 2024)
• How do you eval legal work when experts disagree? A benchmark with one right answer misses most of the job. (Guha et al., 2023)
• What makes a good verifier, human or model? Registry staff using our tools went from filling in checklists to checking them. (Cobbe et al., 2021)
• Which skills are safe to lose to a machine? (Bainbridge, 1983)

anya@jhana.ai · I reply to anything thoughtful.