Selected work
Portfolio crowding, liquidity, and risk investigation
NYIC leadership experience combining streaming portfolio context, deterministic risk services, predictive scores, and historical evidence.
Engineering lessons
How can analysts investigate portfolio risk through controlled AI tools?
- Maintain a shared portfolio context across crowding, liquidity, and risk questions.
- Keep financial calculations in analytical services with inspectable inputs.
- Preserve analyst direction, scoped access, and a clear record of supporting evidence.
Context and professional contribution
Portfolio-risk investigation requires a current view of positions and a broader understanding of how those positions may behave together. Crowding, liquidity, securities borrowing, and shared ownership can affect the interpretation of an exposure. Analysts also need evidence from prior unwind episodes and a reliable way to inspect the calculations behind a recommendation.
This study describes NYIC leadership experience architecting and building a streaming risk platform for an institutional investment environment. The documented work integrated market and portfolio information, analytical services, predictive scores, research retrieval, and an analyst investigation interface. The institution and its positions remain anonymous.
The platform produced cut, keep, and hedge recommendations with supporting evidence. These were analytical outputs for review. Portfolio managers and risk officers retained responsibility for deciding whether the evidence justified a change in positioning.
Assemble a shared portfolio context
The streaming architecture used Kinesis, Apache Flink, Amazon S3, and Neptune Analytics. It combined positions, securities-borrow information, exchange-traded-fund ownership, and market feeds into a shared portfolio-risk context. Governed Snowflake views and Databricks feature pipelines contributed additional analytical inputs.
This common context helped connect measures that originate in different systems. A position record, an ownership observation, and a borrow indicator may use different identifiers and update schedules. The practical requirement is to preserve enough source and time information to understand which observations can be compared.
A streaming design also introduces questions about completeness. An apparent change in risk can result from a real market event, a delayed feed, or an incomplete position update. Freshness checks and observable service behavior are therefore central to the interpretation of the analytical output.
Keep risk calculations in analytical services
Deterministic risk calculations remained in analytical services. Lambda-backed Model Context Protocol tools exposed these services through AgentCore Gateway. Agents using Amazon Bedrock models on AgentCore Runtime combined the calculations with SageMaker XGBoost and CatBoost scores and retrieved historical unwind cases.
This separation preserved a reproducible basis for the numerical analysis. The agent could gather relevant results and explain their implications, while the underlying calculation retained defined inputs and service behavior. Predictive scores contributed an additional analytical signal whose role remained distinct from a measured exposure.
The architecture is especially useful when a recommendation depends on several kinds of evidence. A liquidity measure, a predictive score, and a historical case should not be presented as interchangeable support. A reviewable output identifies what each contributes and what uncertainty remains.
Support analyst-led investigation
A Snowflake Cortex Agent supported portfolio-risk investigation. It combined Cortex Analyst queries over precomputed crowding, liquidity, and exposure measures with Cortex Search retrieval of historical unwind cases. Python and FastAPI integrations connected this capability to the broader platform.
Analysts could investigate risk drivers and inspect the evidence behind the recommendations. The structured query path exposed numerical context, while the search path supplied relevant historical material. This made it possible to move from a recommendation into the analytical details that a reviewer would need to challenge it.
Historical cases can inform a hypothesis without establishing that the current situation will follow the same path. Differences in market structure, positioning, and available liquidity matter. The investigation interface should therefore support comparison and questioning, with each analogy presented in its historical context.
Implement controls across the tool boundary
The documented implementation included role-based tool access, schema validation, data-freshness checks, audit logging, and agent tracing. These controls addressed who could request an analysis, whether the request was well formed, whether its inputs were current enough, and how the resulting execution could be investigated.
Supporting analytical services ran on Amazon EKS. AWS Cloud Development Kit constructs provisioned AgentCore Runtime, Gateway, Lambda integrations, and their identity roles. This connected the workflow to a repeatable infrastructure configuration and explicit access relationships.
The value of these controls is practical. A stale input should produce an inspectable limitation. An unauthorized tool request should fail at the service boundary. A disputed recommendation should have a trace that identifies the calculations, model calls, and retrieved evidence involved.
Evaluation concerns for portfolio support
Evaluation of this class of system should separately examine calculation correctness, source freshness, exposure completeness, recommendation support, and appropriate handling of missing information. These criteria help a risk team assess the behavior of the system under the conditions that matter for its own portfolio and decision process.
Stress scenarios can include interrupted market feeds, delayed position updates, unavailable borrow information, conflicting identifiers, and unusual concentration patterns. Reviewers should also examine whether the explanation remains faithful to the calculated values and whether uncertainty is preserved in the proposed action.
An aggregate accuracy measure can conceal the failure mode that matters most to a risk team. The severity and consequence of a missed exposure or an unsupported recommendation need explicit attention. Portfolio managers and risk officers retain authority over the decision and its implementation.
Documented result and transferable lesson
The delivered platform connected streaming portfolio context, deterministic analytics, predictive scores, historical evidence, and conversational investigation. It made recommendations inspectable through the same governed information and services used to generate them.
The transferable lesson concerns how analytical evidence reaches the investment decision. AI can help organize a complex risk investigation when current numerical evidence, model-derived signals, historical comparisons, and human authority remain clearly connected and individually reviewable.