Activating Alpha Generation With Multi Agent AI for Fund Screening

LSEG's Lipper team refactored its global fund intelligence data into an AI-ready product on the Databricks Platform, centering on semantics, routing, and orchestration rather than clever prompts.

The result is a practical pattern for moving from demos to dependable analytical agents. The core idea: quality comes from the data and the architecture, not from the prompt.

Why naive LLMs struggle with fund data

Fund analytics hide domain complexity inside simple questions. A user asking for lower-fee, higher-performing alternatives to an S&P 500 ETF triggers identifier resolution across markets, taxonomy nuances, long-run performance context, and entity relationships among funds, share classes, benchmarks, and currencies.

Lipper had advantages many datasets lack: robust cross-market identifiers, a rich and hierarchical taxonomy, decades of history across regimes, and entities modeled as professionals think about them. Even so, three issues surfaced in early testing: models slowed and failed when pointed at too many tables, different question types needed different domains, and column-level metadata did not capture how practitioners interpret fields in context.

Engineering the semantic and routing backbone

The team addressed reliability in the architecture. They created purpose-built views that minimize unnecessary joins and scoped each agent's working set. They segmented Genie spaces by data domain, such as profile, performance, and classification, so routing could direct each question to a constrained, well-defined slice of data.

Unity Catalog provides governance, lineage, access controls, and column-level metadata. On top, a semantic layer encodes financial interpretation, taxonomy, standardized joins and filters, and business logic. That layer gives agents consistent definitions and repeatable outcomes. Internally, the same foundation supports natural language fund screening, cross-fund performance comparisons, and experiments with AI-generated fact sheets.

Orchestration that plans, evaluates, and admits uncertainty

Execution is not a simple pipeline. Plans are created, results are evaluated, and the system returns to the same Genie spaces when initial answers are insufficient. The team outgrew an initial Agent Bricks setup and built a custom orchestrator on the Databricks Platform for finer control of planning, evaluation, and re-querying. MLflow traces show iterative LLM calls and tool use until the system is satisfied the answer meets a professional standard.

The orchestrator also flags blind spots. If five-year returns are missing for specific funds, that compromise is surfaced to both developers and end users in the client, such as Claude via a hosted MCP server on the Databricks Platform. Transparency lets analysts decide whether to accept the result or refine the query.

Tune the data, not the model

Most failures that looked like model errors were actually data issues. The fixes lived in the semantic layer and views, not in longer prompts or fine-tuning. Field definitions needed richer meaning, joins needed to explain why they exist, business logic had to be explicit, tricky question patterns needed sample queries, and views sometimes had to be narrowed or broadened.

Stuffing more into prompts created fragility. Investing in semantics, purpose-built views, and lightweight space instructions produced durable gains without model tuning. The strategic bet is straightforward: frontier model labs will keep improving models and context windows, while no one will improve your data for you. The highest leverage is your data foundation.

Reliability that scales with model progress

A data-first design, segmented Genie spaces, and a deliberate orchestrator turned Lipper's deep history into reliable natural language analysis. As foundation models and runtimes get faster, the same architecture will compound in value. Teams that invest in semantics and routing now will see accuracy and speed rise without reworking prompts or retraining models.