Aaron Linsky

Aaron Linsky is Head of Engineering for the Alpha Engine and a partner at Bridgewater Associates. Previously the founding chief technology officer of AIA Labs, he built the technology platform for its AI-first investment research and strategies. He now leads the work of extending and scaling those capabilities across the Alpha Engine.

From investment analysis to engineering leadership

Linsky joined Bridgewater in 2008 as a data analyst on the Equities team before moving into technology. He developed investment logic alongside investors, helped scale the firm’s research platform to the cloud, and led the redesign and unification of the software supporting portfolio operations. He graduated magna cum laude from Princeton University with a degree in Operations Research and Financial Engineering. His career at Bridgewater spans both the investment reasoning that software must express and the infrastructure needed to put it to work.

At AIA Labs, that combination served an ambitious goal: building an artificial investor capable of rigorous, explainable fundamental research. The lab designs systems to reason about the economic causes behind market movements, while also developing tools that help human investors investigate more information and more complex relationships. Its emphasis on explainability has a practical purpose: investors and researchers need to inspect a system’s reasoning, find errors, and improve it.

Linsky’s team began developing Claude-powered capabilities in 2023. Early versions of its Investment Analyst Assistant generated Python code, created data visualizations, and iterated through financial-analysis tasks. In a July 2025 account of that work, he described an assistant that could carry out successive analytical steps within an investor’s workflow.

Forecasting through search and disagreement

Linsky co-authored the November 2025 AIA Forecaster technical report, which describes the team’s system for assigning probabilities to future events using unstructured information such as news. Multiple agents independently research a question, choosing follow-up searches as they learn. A supervisor examines their disagreements and searches for clarifying evidence—for example, checking a disputed fact or base rate—before producing an aggregate forecast. Statistical calibration then adjusts the probabilities to counter the models’ tendency to hedge toward 50 percent.

The report found performance statistically indistinguishable from expert human forecasters on ForecastBench. On MarketLiquid, a harder benchmark drawn from liquid prediction markets, the system trailed market consensus. Yet combining its predictions with market prices improved on consensus alone: the system supplied useful information even where its standalone forecast was weaker. These are results from the collaborative research report, whose authors also caution against extrapolating benchmark performance to other tasks or dispensing with human oversight in high-stakes decisions.

Building research tools investors will use

Linsky’s product-engineering work includes Vault, Bridgewater’s repository and access platform for internal and external unstructured research. It organizes material such as research notes, policy documents, and meeting transcripts for both investors and agents. That shared corpus creates different engineering demands: people need a short list of useful documents, while agents can read more material and generate far more queries.

Two choices illustrate how his team addressed quality and adoption:

• Turn complaints into a test. When investors said Vault search was poor, the team assembled roughly 200 questions and manually identified useful documents, including misleading results that appeared relevant. The evaluation took less than a week. Linsky reported that precision at ten—the share of the top ten results judged right for a question—rose from about 60 percent to roughly 90–95 percent after successive improvements culminating in agentic search. Those figures describe the team’s own evaluation, rather than a general search benchmark.

• Move existing habits before introducing new ones. Better retrieval alone did not make investors open Vault every morning. They already relied on broker-research subscriptions and email folders to keep track of their reading. The team translated subscriptions into saved searches and folders into saved collections. With that familiar organization in place, investors could adopt topic feeds, ask questions across the corpus, and hand collections to agents. Linsky’s approach to adoption treats continuity with existing work as part of the product, alongside retrieval quality.