KelAI

The honest way to read KelAI is as a bet on research orchestration rather than on a model. The company is explicit that it is model-agnostic and applies the latest frontier models to financial research, which means the durable part of the product is supposed to be everything around the model: the Financial Intelligence layer grounded in market data, the Research Orchestration layer that coordinates work, the Data & Integrations layer that reaches your own sources, and the Validation & Evidence layer that keeps an evidence trail attached to every claim. If that orchestration is real, swapping in next quarter's model is a footnote. If it is thin, the product is a chat interface with a finance system prompt. What the site actually documents is deployment shape. Conversational is the entry behaviour — you work directly with an agent to investigate ideas, analyse data, run backtests and produce documented research. Autonomous is the one worth scrutinising: coordinated teams of agents executing complex research workflows continuously. Multi-agent research over financial data is hard in ways that are easy to demo and hard to keep honest, because coordination failures look like confidence. Bespoke is the enterprise escape hatch — wiring KelAI into proprietary data, internal tools and infrastructure so the research layer sits on top of what you already have rather than replacing it. The research loop is the clearest articulation of intent on the site. Investigate moves a hypothesis to structured analysis with a complete evidence trail, described as the way an analyst team would work. Validate tests a thesis against historic data, challenges assumptions and assesses robustness before the evidence informs a decision. Scale runs research across multiple datasets and alpha types in parallel, with capacity that moves up or down. That is a coherent shape for a public-markets desk, and the evidence trail is the part that matters most in practice — research that cannot be reconstructed is research an investment committee cannot use. Governance is stated rather than evidenced. SOC 2-aligned by design against AICPA Trust Services Criteria, isolated workspaces, controlled model access, an on-premises deployment option, and activity logs spanning users, agents and systems. Those are the right controls to name for an institutional buyer, and on-premises plus isolated workspaces is a serious answer to the data-control objection. But these are design principles as written on the site. No audit report is published, no independent verification is visible, and the model-access controls are asserted without describing what the control actually is. The gaps are consistent and worth stating plainly. No underlying AI models are named anywhere, despite the model-agnostic claim — so a procurement team that needs to know which vendors process its data gets nothing from the marketing site. No performance data, no backtest results and no named customers are published. Nothing