AI Street reported two implementation patterns from AI Engineer New York. At Apollo’s Redding Ridge Capital, preparing a proposed portfolio of 200–400 credits previously took two or three days. A Rowspace-based workflow combining document libraries, trade and compliance systems, and current pricing now prepares a portfolio in minutes and gets it roughly 80%–90% complete for portfolio-manager refinement. AI Street
That percentage describes completion of a proposed portfolio. It is not a prediction-accuracy figure and does not establish authority to invest capital. The useful evaluation boundary is the handoff. If a portfolio manager replaces a credit because the market price moved during review, the edit should not automatically be labeled a model error. Preserving the proposal’s prices and information set lets the evaluator separate “the world changed” from “the earlier choice was wrong.”
Coatue’s Frank Long described decomposing research so rules, conventional software, machine learning, agents and people handle different tasks. His example involved Korean-language semiconductor and AI news compared with decades of internal investment notes. The task-routing idea matters because not every step benefits from the same kind of model. The multilingual comparison also makes provenance inspectable: a translation change can look like a company change unless the original wording remains available.
The route through those documents may be anfractuous—full of windings and intricate turnings, pronounced an-FRAK-choo-us. The answer need not be. Merriam-Webster