How AI Turns Technical Analysis from Info‑Heavy to Model‑Driven: A Real‑World Case Study

Experiment: Manual research vs. AI‑assisted analysis

Traditional analysis of a domestic investment bank’s digital‑architecture roadmap required a month of manual work, with roughly 75% of the time spent on gathering blogs, annual reports, technical sites, executive interviews, and GitHub repositories. An AI‑driven notebook was given the same source material and the following prompt:

Goldman Sachs’s digital evolution path, including legacy systems like SecDB. In addition to official content, use related blogs, GitHub, JD, etc., to assist analysis.

Within minutes the LLM produced a concise overview that identified three evolutionary eras:

1.0 Era (1990s‑2010) : Monolithic architecture built around SecDB and the Slang language, providing internal traders with asymmetric information and real‑time risk aggregation.

2.0 Era (2010‑2020) : Platformization and externalization, exemplified by Goldman’s Marquee API‑as‑a‑service model.

3.0 Era (2020‑present) : Ecosystem and cloud‑native phase, featuring the Legend (Alloy) project, open‑source contributions via FINOS, and deep integration with AWS for large‑scale migration.

AI makes explicit knowledge cheap; the remaining bottleneck shifts to model building and hypothesis testing.

Structured modeling – the TARGET framework

Iterative analysis refined an initial “Modern Digital Business” model into a five‑pillar TARGET model that remains the core of technical insight:

Digital Operations : Online, automated processes delivering self‑service and low‑friction delivery.

Platform Strategy & Architecture Evolution : Aligning fintech capabilities with business strategy.

Experience Productization : Building omnichannel digital customer experiences.

Intelligent Decision Mechanisms : Leveraging financial engineering and AI to enhance decision making.

Trusted Infrastructure : Shortening the “idea → usable software” cycle.

AI accelerates knowledge laying, but switching or extending models still requires human judgment.

From SecDB to Athena – modern equivalents

Goldman’s SecDB is a 30‑year‑old system with ~150 million lines of Slang code, ~350 data types, and ~10 000 built‑in functions, compiled with GraalVM to support pricing and risk workloads. Comparable platforms today include:

Bank of America’s Quartz

JPMorgan’s Athena (≈35 M lines of Python in 2018)

These platforms share a three‑layer architecture:

Unified Object Database : Trades, market data, pricing models, and risk metrics are treated as first‑class objects.

Global Dependency Graph / DAG Engine : Automatic recomputation of all impacted risk values when market inputs change.

Highly Integrated Development Environment : Tight coupling of code (Slang or Python) and data, enabling near‑instant propagation of logic changes across the enterprise.

Modern prototyping stack

Building a minimal prototype of an Athena‑ or SecDB‑like system is now feasible with the following mature components:

Jupyter / Python notebooks for interactive development

WebGPU for high‑performance visualizations

DAG and streaming frameworks (e.g., Apache Beam, Dagster)

Open‑source algorithm libraries (NumPy, pandas, SciPy, PyTorch)

Financial‑engineering ecosystems (QuantLib, OpenGamma)

Data‑mesh governance practices (catalogs, lineage, quality checks)

Implicit‑knowledge reasoning – Data Mesh prototype example

To explore the data layer of such platforms, the open‑source visualization tool Perspective from J.P. Morgan was examined. Using an LLM‑enhanced workflow (Cursor + Opus 4.5), a simple data‑mesh prototype was generated, demonstrating how AI can:

Produce scaffold code for data ingestion, cataloging, and streaming.

Guide the selection and configuration of tools (e.g., Kafka for streams, dbt for transformations).

Highlight gaps that require manual validation, such as latency requirements or domain‑specific risk calculations.

Creating a minimal Athena or SecDB‑style prototype is no longer difficult.

Conclusion: Insight shifts from information‑heavy to model‑heavy

Generative AI eliminates the labor‑intensive phase of information gathering, turning the primary constraint from “insufficient data” to “insufficient models.” The new challenge for technical analysts is to construct explanatory structures, formulate stronger hypotheses, and validate implicit knowledge through prototypes and experiments.