A leading Israel-based financial analytics company specializing in investment analysis of publicly traded companies worldwide. The firm offers in-depth evaluation of both global stock markets and individual enterprises using advanced data techniques and proprietary scoring models.
The client’s core value lies in helping institutional investors navigate complex, fast-changing global markets. As their coverage expanded across geographies and sectors, so did the complexity of the data and analysis required.
Every company under coverage had to be evaluated through a wide range of parameters – typically 30 to 40 per enterprise – which varied by industry and sector followed by GICS. For example, real estate entities were assessed based on leverage ratios and collateral lending thresholds, while banks were evaluated on metrics like default rates, interest rate exposure, and inflation sensitivity, etc.
But evaluating individual companies wasn’t enough. The team needed to generate macroeconomic analyses by sector and region – and those analyses had to reflect up-to-date conditions across major economies, including the U.S., Europe, Asia, and emerging markets like Brazil and Thailand.
At scale, manually researching and writing this analysis was unsustainable. The company needed a way to automate macroeconomic narratives, keep them current, and ensure they aligned with the nuanced demands of different sectors and geographies.
The client faced a complex challenge: consolidating diverse, fragmented financial and macroeconomic data to produce actionable investment insights across multiple sectors and geographic regions.
To achieve their goal, the company also had to navigate several significant operational and technical constraints:
To meet these challenges, the team designed an AI-powered data and content pipeline that merged domain-specific LLM usage with live search integration and sector-based logic.
1. Multi-Level Sector Modeling and Parameterization
Using Python, PySpark, and AWS Glue, the team built a processing layer that structured companies according to their GICS sectors. Each sector came with its own set of evaluation metrics – typically 7-8 key economic indicators relevant for macro analysis. For example:
This logic ensured that every company could be analyzed in context – not only by its financials but by the sectoral trends shaping its environment.
2. Region-Specific Macro Analysis Framework
The pipeline then extended this logic to regional macroeconomics. For each sector, macro narratives were generated across multiple layers:
This hierarchy allowed the team to match every company with the most appropriate macroeconomic context. For example, a U.S.-based bank would receive analysis based on the latest American inflation figures and Federal Reserve policies, while a Japanese manufacturer would be aligned with Asia-Pacific trends.
3. LLM Text Generation with Live Internet Context
To overcome the limitations of static LLM knowledge, the team integrated Google Search Tool directly into the Gemini pipeline. When prompted to produce macroeconomic text, the LLM first triggered a series of targeted online searches – for instance:
The search results were used as real-time context for the generation process. Gemini no longer relied on outdated training data. Instead, it wrote based on the latest information gathered moments earlier.
To improve relevance and quality, the system was instructed to prioritize high-authority sources with wide citation reach, such as Bloomberg, CNN, and CNBC. This flexible filtering method worked better than hardcoded source lists.
4. Experimentation, Monitoring, and Tooling
All experiments and outputs were tracked using LangChain, LangFuse, and Deepnote. These tools allowed the technical team to monitor prompt behavior, track changes in generation logic, and rapidly iterate on performance.
5. Final Company-Macro Matching Engine
At the final stage, each company was matched with the best available macroeconomic analysis. The matching logic worked in layers – starting from country-level (if available) and falling back to regional insights when country-specific data was missing.
The resulting profiles combined:
This blend ensured that the final output was both numerically grounded and economically current.
75%+ Reduction in Manual Macro Research Time
Replacing analyst-driven macro reviews with automated generation validated against trusted sources.
Dynamic Global Market Coverage
Macro insights were now available for 7+ regions and 10+ countries, allowing clients to evaluate companies in new emerging markets with updated context.
Consistent and Transparent Investment Scoring
Every company was evaluated with both micro-level financials and macroeconomic overlays, enabling better buy/sell decisions.
Production-Ready LLM Stack
Delivered a full-stack generative AI pipeline using Gemini + Google Search Tool, with modular components deployable across other client segments or analysis types.
Accelerated reaction to market signals
Automatically re-analyzes companies when key metrics (e.g., cash flow, internal scores) shift, keeping reports current without manual intervention. Faster turnaround on updated investment recommendations and reduced lag between data changes and insight delivery.
Reduced operational burden across 9,000+ companies
Replaced repetitive, manual company reviews with an automated system that processes 100-120 financial and sector-specific parameters per entity.
