On 69 active U.S. equity funds, AI personas built from disclosures reproduced manager calls with macro-F1 0.410 vs 0.334 and were preferred over a generic advisor in 69% of judged comparisons.

LLM summaries of MD&A sections and earnings calls flipped the bull or bear call 24 to 32% of the time; source-checked summary selection cut flips by 20 to 38%.

Six LLMs showed consistent sector, size and contrarian preferences and held on to them even against counter-evidence.

26K expert-annotated examples testing whether LLMs pick the right filing, then the right passage, across S&P 500 companies.

5,703 expert-annotated query, evidence and answer triplets from 10-K filings, written the terse, jargon-heavy way analysts actually ask.

An extraction agent and a text-to-SQL agent turn KPIs and guidance in filings and calls into queryable data.

Our embedding model ranked #1 for retrieval on the MTEB leaderboard (score 60.2, May 2024).

Trains multilingual embeddings to match a monolingual teacher, beating hard-label contrastive learning on bitext mining and semantic similarity.

Retrieve evidence, have the model explain its own errors, then revise; fewer factual errors at lower token cost than prior methods.

Eval-RAG grades LLM answers against retrieved documents and agrees more closely with expert judgments than other evaluators.