EP. 44 - She Quit at 29 With No Code, No Co-Founder, No Funding – Now Hudson Labs Is Finance AI that has a “No Hallucination Guarantee”

She Quit at 29 With No Code, No Co-Founder, No Funding – Now Hudson Labs Is Finance AI that has a “No Hallucination Guarantee”Kris Bennatti (@)  CEO Hudson Labs (Toronto), a former forensic-accounting data scientist who built the first LLM for finance, back in 2019 before anyone knew what to make of it. Her co-founder/CTO Suhas Pai literally wrote the book – O'Reilly's "Designing Large Language Model Applications." "If you try to pull multi-period KPIs beyond four quarters with a generalist tool, you'll get one wrong number 30% of the time."We cover:The leap: quitting a great job at 29 with no funding, no co-founder, and not one line of code written  "those choices are just a reflection of poor risk assessment capabilities"The first hard lesson: the academic papers claiming AI could predict fraud were poisoned by target leakage, the training data contained the outcome. The product she quit her job to build didn't work, and she rebuilt from scratchThe forensic risk score: 70+ means a one-in-three chance of an SEC enforcement action within three years and 3x the likelihood of a securities lawsuit  used by investors, D&O insurers pricing risk, and plaintiff-side class-action firms. "We specialize in selling to the enemy."Why generalist LLMs hallucinate on financial numbers: effective context length vs the advertised window, the two-to-three-earnings-calls attention limit, and why hallucination is really LLM memory bleeding into real dataThe Hudson Labs answer: AI-specific pre-processing of every filing, transcript, and presentation retrieve the whole table, the units, the currency, consolidated-vs-segment and a no-hallucination guarantee, the first of its kindTone as a screen: find the most stressed-out CEOs, the most confident CFOs, track deflection on analyst Q&A over time  impossible unless tone is stored in the embeddingsThe soft-guidance problem: why "capex is now expected to be…" slips past generalist AI, and the pathways built to never miss a guidance cueThe server useful-life screen: every hyperscaler raising estimates in a follow-the-leader wave and Amazon as the only one to cutThe cost story nobody's pricing: a 500-company AI-ecosystem deep dive that cost $200 on Hudson Labs vs ~$13,000 on a frontier-model API at lower accuracy  and why cost efficiency becomes the tailwind as subsidies fadeThe business today: ~100 hedge fund clients plus Fortune 100s, insurers, and law firms and a new $100/month tierHighlights:  (1:13) Quitting at 29 with nothing but the problem (5:34) The target-leakage discovery — the papers were wrong (9:46) Suhas Pai — the co-founder who wrote the book (13:10) How Hudson Labs differs from the AI-startup wave (15:30) Why generalist LLMs get numbers wrong — context and attention (23:24) The worst strategic decision — and pivoting when short sellers shrank (28:21) Demo: KPIs with verbatim call commentary, no hallucinations (35:30) The infrastructure behind the guarantee (39:34) Tracking Waymo vs Uber when the stats aren't standard (46:39) The forensic risk score — and Meta's rising off-balance-sheet risk (52:00) The server useful-life wave — everyone up, Amazon down (55:14) $200 vs $13,000 — the deep-dive cost math (57:00) The next three years and the new $100/month tier______________________________________________________________________💡 This episode is powered by:Fiscal.AI: Contact SalesPitch The PM Links:📩 Subscribe to our Substack for research updates and new high-conviction episodes from top PMs, and our Job Board: ⁠https://pitchthepm.substack.com/⁠Doug Garber on LinkedIn for daily market color: : https://www.linkedin.com/in/doug-garber-42aa508 Hudson Labs Links:Kris Bennatti: https://www.linkedin.com/in/kbennatti/ Hudson Labs: https://www.hudson-labs.com/