Selected work
Macro-event research with point-in-time evidence
NYIC leadership experience connecting economic releases, historical analogs, pricing tools, and portfolio-aware scenario generation.
Engineering lessons
How can macro research remain reproducible as information changes?
- Preserve the information vintage attached to an event and each analytical input.
- Coordinate evidence collection, pricing, and risk work through explicit workflow state.
- Retain the sources and calculation context behind the research output.
Context and professional contribution
Macro research connects economic events with market reactions and portfolio implications. A central-bank announcement or economic release can require analysts to compare expectations, revised figures, historical analogs, current pricing, and existing exposures. The work is especially sensitive to timing because later information can make an earlier decision appear easier than it actually was.
This study describes NYIC leadership experience designing and building a production agent framework for rates, foreign exchange, and credit analysis. The documented implementation connected economic releases, central-bank events, historical analogs, live market information, and portfolio state through Python and SQL services.
The application produced structured scenarios and candidate-trade outputs with supporting evidence and portfolio constraints. Its analytical role was to organize research and scenario preparation. An output describing a possible action did not itself constitute an executed trade.
Preserve the event and its information vintage
The architecture modeled release vintages and event timestamps in Snowflake Cortex Analyst semantic views. These views covered economic releases, historical market reactions, and portfolio exposures. Verified SQL examples supported the queries used to retrieve historical analogs and assemble scenario inputs.
Release vintages are essential for point-in-time reasoning. An economic figure may be revised after publication, while market prices and portfolio exposures continue to change. Keeping the original release and later revisions distinguishable allows an analyst to ask what information was available at the moment under review.
The same discipline applies to historical analogs. A superficially similar event may have occurred under different market conditions or policy expectations. Structured inputs and source references help reviewers examine whether the comparison is meaningful for the current question.
Coordinate research with stateful workflows
Stateful LangGraph workflows ran on Amazon Bedrock AgentCore Runtime. Hybrid retrieval connected OpenSearch, Snowflake, and Databricks. Retrieved evidence, source references, and portfolio constraints traveled with the workflow into the scenario output.
Stateful orchestration matters because a macro analysis may require several dependent steps. The system can gather an event record, inspect comparable episodes, query current exposures, and evaluate a candidate scenario. Each step needs enough context to preserve the analytical question and the relevant time boundary.
The design also used conditional routing and bounded retries to handle missing or stale information. A workflow should respond differently to a temporarily unavailable pricing service and an event whose required release detail is absent. Keeping these conditions explicit supports a more useful handoff to an analyst.
Keep pricing and risk calculations reproducible
Typed Model Context Protocol interfaces exposed pricing, risk, research, and portfolio APIs through AgentCore Gateway. Pricing and risk calculations remained in deterministic services. The model could coordinate the analysis and explain the results while the numerical operations retained their own reproducible implementation.
This boundary is particularly important when a scenario changes several assumptions. A reviewer needs to know which inputs drove the calculated effect and whether the relevant portfolio constraints were applied. A narrative alone cannot establish that a pricing or exposure calculation was performed correctly.
The tool contract provides a place to validate identifiers, units, dates, and parameter ranges. It also allows the analytical service to return an explicit failure when the requested calculation is unsupported or its inputs are incomplete. The agent can then preserve that limitation in the scenario discussion.
Make evidence useful in the final research output
The framework carried source references into structured scenarios and candidate actions. This linked the final interpretation to the economic information, historical comparisons, and portfolio state used during the workflow. An analyst could examine the evidence and decide which assumptions deserved further investigation.
Scenario generation remains conditional reasoning. A scenario describes an outcome under stated assumptions, with uncertainty about whether those assumptions will hold. The architecture is most useful when it preserves alternatives and identifies which observations would change the interpretation.
Portfolio awareness adds another constraint. The same macro view can have different implications for portfolios with different existing exposures and limits. The documented workflow carried those constraints with its analytical output so reviewers could assess its relevance to the portfolio.
Deployment and evaluation
Reusable AWS Cloud Development Kit constructs provisioned AgentCore Runtime, Gateway, access roles, and supporting resources. Containerized analytical services ran on Amazon EKS through continuous delivery. This connected the agent layer to deployable analytical services with explicit infrastructure ownership.
The documented evaluation covered retrieval, tool execution, routing, and scenario outputs, including preservation of source references and portfolio constraints. These checks addressed the system trajectory as well as the final narrative. They provided a basis for investigating whether a failure arose from evidence selection, tool behavior, or synthesis.
Additional review criteria for this class of system include point-in-time validity, numeric consistency, appropriate analog selection, and handling of contradictory indicators. These criteria help reviewers determine whether a scenario preserves the temporal and analytical conditions needed for an informed investment discussion.
Documented result and transferable lesson
The implemented framework connected event information, historical research, pricing and risk tools, and portfolio context in a production research workflow. The architecture made the research trajectory inspectable across its retrieval, calculation, and synthesis stages while retaining the institution’s information boundaries.
The central lesson is that macro AI requires disciplined temporal context. A persuasive scenario is useful only when the evidence, release vintage, calculation inputs, and portfolio constraints remain available for review. Human investment judgment determines whether the scenario warrants action and how its uncertainty should influence that decision.