← All speakers

Vinoo Ganesh

Conference affiliation: CEO & Co-Founder · Kepler · 2026

Vinoo Ganesh is CEO and co-founder of Kepler, which builds verifiable AI for financial services and other work that requires traceable evidence. His defining design choice separates proposing an analysis from producing its numbers: language models plan the work, while conventional software retrieves stored facts, performs calculations, and connects the results to their sources. The aim is to help investment teams produce models and memoranda they can inspect and defend.

On this page

From data infrastructure to investment research

Vinoo spent seven years at Palantir working across infrastructure and customer deployments. His responsibilities included search and indexing systems, the Spark/YARN compute platform, and engagements spanning commercial, healthcare, and defense customers. He also ran Project Frontline, Palantir’s rotation program for putting engineers into forward-deployed roles.

After Palantir, he became chief technology officer of Veraset, a geospatial data startup that was subsequently acquired. His work migrating data products from CSV to Apache Parquet examined how a storage-format decision affects data quality, delivery costs, and customers’ workflows. He later led business engineering for Ashler at Citadel, overseeing investment platforms, data pipelines, and the teams supporting them.

At Citadel, Vinoo saw portfolio managers combine investment judgment with deterministic tools. That supplies a concrete model for his approach to AI: an analyst’s ability to decide what to investigate depends on instruments that retrieve and calculate accurately. In his explanation of Kepler’s architecture, he applies the same division of responsibility to language models.

Vinoo founded Kepler with John McRaven, another former Palantir engineer. Their founding vision centers on work products such as investment memoranda, discounted cash-flow models, and fairness opinions. These require more than plausible prose: readers need to examine the figures and assumptions behind a conclusion.

His writing on forward-deployed engineering also develops a product strategy from his deployment experience. Engineers embedded with customers observe workflows, ship narrowly scoped solutions, and use those discoveries to improve the shared product. He applies that approach at Kepler.

What a citation must prove

Vinoo distinguishes three numerical failures: retrieving the wrong document, selecting the wrong figure from a relevant document, and calculating incorrectly from the right inputs. His annual-report example makes the middle failure clear. A company’s 2025 filing can contain its 2023 revenue. Linking the correct filing does not establish that the answer used the correct year’s column—or preserved the table’s currency and scale.

He also describes a financial-extraction model that Kepler’s team trained to reach 94 percent accuracy. His objection is practical: an aggregate score offers little protection to an investor using one of the incorrect figures. He argues that evaluations of a probabilistic model cannot substitute for checking the particular facts and calculations delivered to a user.

Three principles organize Kepler’s response:

  • Atomic provenance: The model refers to a registered fact rather than supplying the displayed value itself. Software retrieves that value with its source and context, including unit, scale, and reporting period. A subsequent check removes figures that cannot be verified against the stored evidence.
  • Scope determinism: The model decides what needs calculating; predetermined code performs the arithmetic. A question about adjusted earnings may require finding filings, considering later disclosures, and identifying relevant adjustments. Those planning decisions remain separate from retrieving values and computing the result. Vinoo also treats this as a cost decision: ordinary arithmetic does not warrant an expensive model call.
  • Derivation chains: A calculated figure retains its inputs, formula, and applicable definitions. Two investment teams may treat recurring expenses or enterprise value differently. Verification establishes whether a result follows the firm’s stated rules; it does not establish that its investment thesis is correct.

This last distinction preserves room for disagreement. Two investors can inspect the same evidence and take opposite positions. Vinoo’s proposed infrastructure makes their calculations and assumptions inspectable without claiming to resolve the investment decision for them.

Choosing restraint over a more impressive answer

Vinoo accepts a product tradeoff that can make an assistant appear less capable: withhold unverifiable numbers. In his technical account of Kepler’s citation system, he explicitly distinguishes the model from the platform. The model can still hallucinate. The platform checks references against a citation registry, attempts repair in selected cases, and strips numerical claims it cannot substantiate. Those checks are part of the architecture, rather than a source list attached after an answer has been written.

Retrieval remains a separate problem. Vinoo describes combining keyword and similarity search, then reranking results against the question and the trustworthiness of their sources. A filing and an earnings-call transcript may report conflicting information; finding both does not remove the need to resolve which disclosure applies. Private research adds another requirement: he describes reauthorizing document hits in Postgres so that even two desks within the same institution do not inadvertently gain access to one another’s work.

Kepler’s product carries this approach into research, workbooks, and firm-specific financial definitions. Its workbooks preserve formulas and sources when exported to Excel. Vinoo describes the immediate value as returning analysts’ time: gathering filings and preparing an initial financial model should consume less of their day, leaving them more time to investigate and interpret.

His writing on reliable AI context extends these concerns to lineage, versioning, validation, reproducibility, and observability. He also teaches data and AI engineering, advises startups, and serves on Washington University’s McKelvey Engineering National Council.

Vinoo sees applications beyond finance, including checking legal precedents against stored cases and tracing scientific claims to source material. He presents these as extensions of the architecture. His narrower promise for investment work remains important: traceable research and correct arithmetic can support professional judgment without amounting to a complete replacement for investment professionals.

Read the topics behind these talks

4 conference talks

Loading more talks and highlights…

References