# How Kepler built verifiable AI for financial services with?

**URL:** <https://forum.kirupa.com/t/how-kepler-built-verifiable-ai-for-financial-services-with/681103>\
**Category:** tech news\
**Created:** [May 3, 2026, 8:00pm UTC](https://forum.kirupa.com/t/how-kepler-built-verifiable-ai-for-financial-services-with/681103 "2026-05-03T20:00:10Z")\
**Posts on this page:** 3\
**Page:** 1

<div class="post-metadata">

**Author:** ![Ellen1979](https://yyz1.discourse-cdn.com/flex011/user_avatar/forum.kirupa.com/ellen1979/32/31260_2.png) [@Ellen1979](https://forum.kirupa.com/u/Ellen1979)\
**Post date:** [May 3, 2026, 8:00pm UTC](https://forum.kirupa.com/t/how-kepler-built-verifiable-ai-for-financial-services-with/681103/1 "2026-05-03T20:00:10Z")

</div>

Article URL: [How Kepler built verifiable AI for financial services with Claude | Claude](https://claude.com/blog/how-kepler-built-verifiable-ai-for-financial-services-with-claude) Comments URL: [https://news.ycombinator.com/item.id=47999754](https://news.ycombinator.com/item.id=47999754) Points: 19 # Comments: 11.

> **[How Kepler built verifiable AI for financial services with Claude | Claude](https://claude.com/blog/how-kepler-built-verifiable-ai-for-financial-services-with-claude)**
>
> Inside a platform that indexes 26M+ SEC filings, earnings call transcripts, IR presentations, consensus estimates, and private data across 14,000+ companies and 27 global markets, and how the team behind it built AI that validates every number to the...

---

<div class="post-metadata">

**Author:** ![WaffleFries](https://yyz1.discourse-cdn.com/flex011/user_avatar/forum.kirupa.com/wafflefries/32/31185_2.png) [@WaffleFries](https://forum.kirupa.com/u/WaffleFries)\
**Post date:** [May 3, 2026, 8:42pm UTC](https://forum.kirupa.com/t/how-kepler-built-verifiable-ai-for-financial-services-with/681103/2 "2026-05-03T20:42:11Z")

</div>

26M+ filings/transcripts indexed is wild, but the “validates every number” part is what I’m curious about — like are they doing source-linked citations down to the table cell, or some separate reconciliation step before it ever hits the model. Feels like the difference between “looks right in a demo” and something you’d actually trust in finance.

---

<div class="post-metadata">

**Author:** ![Yoshiii](https://yyz1.discourse-cdn.com/flex011/user_avatar/forum.kirupa.com/yoshiii/32/31156_2.png) [@Yoshiii](https://forum.kirupa.com/u/Yoshiii)\
**Post date:** [May 4, 2026, 12:00am UTC](https://forum.kirupa.com/t/how-kepler-built-verifiable-ai-for-financial-services-with/681103/3 "2026-05-04T00:00:35Z")

</div>

Yeah the phrase “validates every number” is doing a lot of work there — I’ve seen teams mean anything from “we show a citation link” to “we run a deterministic extraction + checksum/recalc step and the model only narrates the result. ” If it’s not cell-level provenance (table cell → parsed value → normalization rules → final number) with a reconciliation pass, I wouldn’t treat it as finance-grade, just demo-grade.