Kepler

Where did that number come from? A link to a filing doesn't prove you pulled the right value from it. In this World's Fair 2026 recording, Vinoo Ganesh, CEO and cofounder of Kepler, walks through how they stop models from inventing financial figures. The model picks a reference. Software retrieves the value, checks it against the source and runs the arithmetic. A record of the inputs and calculations lets you work backward from the result. Watch "How Kepler Built Verifiable AI for Financial Services": https://lnkd.in/eKbSr9mE Vinoo joins us in New York on October 13 with "Do Not Let Models Invent Numbers," an Expo Session included with Engineering admission. AI Engineer New York, October 12 to 14. Tickets: https://ai.engineer/nyc

Pipeline changes at Kepler used to take 1–2 weeks to ship confidently. With Prefect, they take 1–2 days. Kepler replaced its homegrown orchestration stack in four weeks and has processed over 20 million documents through Prefect. Orchestration maintenance is close to zero, down from fixing problems about twice a week. “We’ve leveraged Prefect’s primitives to make development a breeze.” — Sean Gorsky, founding engineer at Kepler

Socrates thought that writing was overrated. We know this because Plato wrote it down. Last week a philosophy professor made the same case in The New York Times. Brave, given the choice of medium. I reflect on my frustrations with this professor in Kepler's latest blog piece. It has since become mandatory reading for our new engineers (or at least I am trying very hard to convince the team it should be). For anyone building with AI, or with otherwise strong opinions on the matter, link is in the comments.

Today's list of open roles at breakout startups and early-stage founders looking for team members, plus interesting moves lately in tech / AI: - Descript: hiring a Brand Design Lead for up to $185K - Shield AI: hiring a Head of Product Design for up to $420K - OP Labs PBC: hiring a Partnerships Lead, Strategic Accounts for up to $299K - Extend: hiring a Startup / Field Marketing Lead for up to $180K - Kepler: hiring a Founding Marketer ...and more.

All data drifts. Fields get renamed, merged and split, and almost nobody writes down how. Not even in SEC filings: 194 of 200 public companies reshaped their financial statements, and none published a map of what changed. A smarter model can't read a map that doesn't exist. Christopher Martin on our team just published a paper that rebuilds the map from the filings themselves. With math, not AI. All 200 companies in a quarter of a second. Chris is one of the best engineers I've worked with. He spent 14 years at Palantir leading optimization after a PhD in aerospace engineering. In the video he puts it simply: "You can't rely on the process that gave you the answer to trust it." That's why we split the work at Kepler: AI where it's strong, math and code where it isn't. It's the same work we do on every firm's own data. This is the depth of research that differentiates a real AI infrastructure platform from a wrapper. And we are only getting started. Link in comments.

I spoke at GAIIM at Columbia yesterday. The question that kept coming up: is AI going to replace the PM? Plenty of people there pitched a vision of yes, eventually. We strongly disagree. Alpha comes from people: what they look for, what they ignore, when they act. As LLMs become commoditized, they do not by themselves provide this alpha. However, AI will still be a massive differentiator for the firms that point it at the right work. The hard part is deciding which job goes to which. The doomsayers aim AI at replacing the thinking: form the view, pick the names, make the call. Meanwhile, the manual work underneath is still mostly manual and a ripe target for automation. Pulling numbers out of 10-Ks. Checking whether fiscal Q3 means calendar Q3. Catching the figure that got restated. Re-checking every number before it reaches the PM. My talk was about flipping that. Every firm needs the same foundation: data connected, documents read, definitions pinned down, every number pulled and computed by code and traced to its source. Buy that. Build your own workflows and agents on top of it, because that's where your process lives. Then your analysts and PMs spend their time on the thinking and the decision - where the alpha lives. That foundation is what we build at Kepler. Thank you Ehsan Ehsani for having us.

Every upgrade this year made our models smarter, and after every upgrade we gave them less to do. The better they got, the harder their mistakes were to see. A language model doesn't copy a number. It retypes it, and it's a very good typist. The mistakes that survive are the ones that look right. AI output needs a provable provenance layer: that's Kepler. Hallucination is only part of the problem. Full write-up in comments.