Why LLMs Fail at Tax Accuracy

What if the biggest failure of AI today isn't creativity - but math?On this episode of the *1st10 Podcast*, host Boris Epstein sits down with Sreerama Tripuramallu, a repeat founder whose career spans NVIDIA, a Sequoia-backed startup exit to Harvey AI, and now a stealth consumer product tackling one of the most painful problems professionals face: taxes.Sree walks through his unconventional career path - from being the person whose personal-finance post is still pinned inside NVIDIA's Slack, to building Mirage through the pre-ChatGPT era, pivoting fast during the AI boom, and ultimately exiting just two years in. And now, after seeing first-hand how broken tax planning is, Sree makes a contrarian claim: general-purpose LLMs will never be good at taxes. Not because they aren't smart, but because accuracy and precision matter more than language.The discussion dives into a myriad of topics, such asThe Hidden Cost of General AI: Large language models achieve only 27-33% accuracy on tax computations, revealing a massive gap between general AI capabilities and domain-specific precision.The Acqui-Hire Formula: Getting acquired is sometimes about building relationships with shared investors and finding cultural fit with fast-growing companies at the right moment.The Impending CPA Crisis: 75% of CPAs will retire in the next 10-15 years with no clear succession plan.The User-First Building Philosophy: Founders who optimize for end-user experience over technology choices automatically make better trade-offs. The Serendipity of Startup Inception: The best startups often emerge from problems founders can't stop thinking about, combined with unexpected investor interest.Don't miss the part where Sree explains why he believes we're in an AI bubble that won't actually pop, while making a clear comparison to the dot-com boom (and bust) of early-2000s!Chapters00:00 Introductions02:01 From Personal Finance Guru to NVIDIA Engineer06:06 When "Stop Trying" Worked Better Than Hustling10:12 Building Side Projects Until Burnout Forced a Choice14:05 NVIDIA → Mirage → Harvey20:23 Finding Community Inside a Rocketship26:51 A Side Obsession Turns Serious32:17 A Very Different Kind of AI Product37:25 Careers @ Sree's Stealth Startup: Applied AI Engineers42:41 AI Won't Kill Jobs And The AI Bubble WON'T Pop46:05 Conclusion & Contact Details

Quotes:"I stopped worrying about the job and I started worrying about enjoying what I was doing. And it's kind of how I approached everything in my career." - Sreerama Tripuramallu (07:25)"It's not that I'm against paying tax - I just want control over how it's paid." - Sreerama Tripuramallu (30:31)"The LLMs have anywhere from like a 27 to 33% hit rate when it comes to tax computation... My hypothesis from the very beginning was that these models are not going to be able to solve tax." - Sreerama Tripuramallu (34:40)"Excel didn't take away a bunch of CPAs and accountants. It created more of them. [AI is] going to be transformative for all professional work, but it's going to change how we do work." - Sreerama Tripuramallu (43:08)Follow:Spotify: https://podcasters.spotify.com/pod/show/1st10podcast Amazon: https://music.amazon.com/podcasts/7e8ec9af-f38c-4cd9-8c68-1c1dd4516b 27/1st10-podcastApple: https://podcasts.apple.com/us/podcast/1st10-podcast/id1760411207 Podcast: https://www.1st10.com/podcast RSS: https://anchor.fm/s/f951319c/podcast/rss YouTube: https://www.youtube.com/@1st10podcast Connect with usWebsite: www.1st10.comPodcast: www.1st10.com/podcast Twitter www.x.com/1st10engineersLinkedIn: www.linkedin.com/company/1st10/YouTube: www.youtube.com/@1st10podcast Links:Sree's LinkedIn: https://www.linkedin.com/in/sree-tripuramallu/Mirage - acquired by Harvey AIHarvey AI: https://www.harvey.aiSequoia Arc: https://www.sequoiacap.com/arcMusic by Roman Senyk from PixabayProducer: Shrikant Joshi