Selected work
Earnings analysis across evidence and predictive signals
NYIC leadership experience integrating earnings audio, documents, market information, and specialized research agents.
Engineering lessons
How can an earnings event be analyzed across different forms of evidence?
- Align audio, documents, market context, and model inputs around the same event.
- Expose predictive and numerical tools through explicit interfaces.
- Keep analytical roles, source grounding, and evaluation boundaries visible in the final research.
Context and professional contribution
An earnings event produces several accounts of a company’s performance at once. The financial filing, presentation, spoken discussion, and market response each contribute different evidence. Research teams also need to connect the event with portfolio exposures and predictive signals while the information is still operationally relevant.
This study describes NYIC leadership experience designing and implementing an institutional multimodal earnings-analysis platform. The documented work covered event ingestion, model-inference integrations, multi-agent orchestration, numerical research tools, deployment, and operational monitoring. The public account uses the AWS implementation narrative and keeps the institution anonymous.
The platform brought these inputs into structured earnings briefs and sizing inputs with supporting sources. Its purpose was to organize an evidence-intensive research process and make the relationship between analytical viewpoints, numerical results, and source material inspectable.
Align evidence around the event
An event-driven information layer used Kinesis, Amazon S3, and Lake Formation for call audio, transcripts, filings, presentations, and market data. Governed datasets were organized by issuer, event, and timestamp. These identifiers provided the common context for retrieval and predictive model inputs.
This organization addresses a subtle source of error. A passage can be relevant to a company while belonging to a different reporting period or event. A financial table can use a different basis from the spoken discussion. Retaining event and time context gives downstream tools the information needed to avoid combining incompatible evidence.
The architecture preserved several modalities within the same research workflow. Audio and transcripts supported analysis of the call, while filings and presentations supplied documentary context. Market information and portfolio records added the operating context needed to interpret the event for a specific research task.
Connect predictive models through explicit tools
Databricks feature pipelines connected to SageMaker models. XGBoost, CatBoost, and temporal-model predictions were exposed through model-inference tools used by the agent workflow. This kept predictive outputs available as structured analytical inputs with a defined service boundary.
The distinction between a prediction and a reported fact remained important. A predicted response can inform a research hypothesis, while a filing passage supports a statement about reported performance. A well-designed brief should identify which kind of evidence underlies each conclusion and preserve the uncertainty associated with model outputs.
Tool interfaces also make the predictive layer replaceable. A research workflow can consume a stable result contract while the underlying model evolves through its own evaluation and release process. That reduces unnecessary coupling between agent orchestration and the training implementation.
Examine the event through multiple analytical roles
The implementation used bull, bear, forensic, and risk agents in LangGraph with Amazon Bedrock models. AgentCore Runtime hosted the orchestration. Retrieval, market-data, portfolio-exposure, and model-inference services were exposed as controlled Model Context Protocol tools through AgentCore Gateway.
The specialized roles encouraged examination of different aspects of the event. Their outputs were combined with predictions and documentary evidence to produce structured briefs and sizing inputs. The architecture made each role’s supporting material available for review, allowing an analyst to examine the basis of agreement or disagreement.
Multiple agents can still share the same blind spot if they receive incomplete or incorrect inputs. The practical lesson is to evaluate evidence diversity, numerical consistency, and the reconciliation process alongside the individual role prompts. Disagreement can be useful when its source is clear and the final output preserves the unresolved question.
Ground numerical research in governed queries
Python tools called Snowflake Cortex Analyst over semantic views of earnings metrics and portfolio exposures. Generated SQL executed in Snowflake, while Cortex Search retrieved supporting transcript and filing passages. Issuer and event filters constrained the information supplied to the four agents.
This combination connected structured financial questions with unstructured documentary support. A user could inspect a numerical result and the source context needed to interpret it. The semantic layer supplied agreed definitions for the analytical query, including the financial concepts that shape the interpretation of its results.
For this pattern, query correctness and source support are distinct evaluation concerns. A valid query can answer the wrong period, and a relevant passage can fail to support a specific numeric assertion. Both need to be checked within the end-to-end research task.
Deploy and observe the complete workflow
Dockerized Python and FastAPI services and workers ran on Amazon EKS through Terraform and continuous delivery. Reusable AWS Cloud Development Kit constructs provisioned AgentCore Runtime, Gateway, tool adapters, and their identity roles. The documented design kept infrastructure ownership for EKS and AgentCore resources separate.
Regression checks covered tool calls, source attribution, and agent completion. CloudWatch and X-Ray instrumentation tracked latency, token usage, and service failures across agents and supporting services. This operational visibility helped connect an incomplete brief to the underlying model, tool, or service behavior.
Documented result and evaluation boundaries
The implementation produced structured earnings research supported by event-aligned evidence, model outputs, and governed numerical tools. Its operational value rests on the ability to examine the event through several analytical perspectives while keeping the evidence and calculations available for review.
Future evaluation should examine event attribution, numerical fidelity, source coverage, conflicting evidence, and the usefulness of the brief to an authorized analyst. Trade sizing and execution remain subject to institutional review. The transferable lesson is that multimodal intelligence becomes useful through disciplined evidence alignment and reliable analytical interfaces, especially when several sources describe the same event differently.