Selected work

An evidence-grounded security investigation workspace

NYIC leadership experience unifying security alerts, deterministic enrichment tools, source-linked assistance, and analyst approvals.

Engineering lessons

How can an investigation assistant assemble evidence across security systems?

  • Build a case from deterministic operational queries and approved knowledge sources.
  • Preserve evidence context and institutional information boundaries.
  • Judge assistance through the investigation workflow and the analyst’s ability to verify it.

Context and professional contribution

A security investigation often begins with a narrow alert and expands across identity, endpoint, application, entitlement, and vulnerability information. Analysts need to assemble that context quickly while keeping sensitive business material within its approved boundary. The resulting decision may involve a ticket update, an escalation, or a response action with operational consequences.

This study describes NYIC leadership experience designing and shipping an internal investigation application for teams supporting investment-banking systems. The documented role covered backend services, an analyst workspace, retrieval-grounded assistance, deterministic integrations, information controls, evaluation, and phased rollout.

The public account keeps the institution and its security environment anonymous. The application assisted the investigation workflow, while analysts retained approval over ticket updates, escalations, and response actions. That authority boundary was part of the implemented design.

Assemble the case across security systems

The platform brought security information and event management, endpoint, data-loss-prevention, identity, and application alerts into one case workflow. Containerized Python and FastAPI services supported a React and TypeScript analyst workspace. Asynchronous workers and durable messaging enriched alerts with identity, asset, entitlement, and vulnerability information.

This integration addressed the fragmented nature of an investigation. An alert alone can identify an event without establishing its relevance to a specific asset or user. Enrichment adds the context needed to formulate useful questions and decide which evidence should be examined next.

Durable messaging supported work that could not complete within one synchronous request. The analyst interface could reflect the state of the investigation as evidence arrived, while backend services handled the integration work. This helped keep the case workflow connected to the progress of its dependent services.

Ground assistance in approved operational knowledge

A retrieval-grounded assistant used approved runbooks, asset inventories, and resolved cases. It returned source-linked summaries, identified missing evidence, and suggested next steps for analyst review. The source boundary kept the assistant focused on material approved for the investigation task.

Runbooks and prior cases serve different purposes. A runbook describes an approved procedure, while a prior case offers an example whose relevance still needs judgment. An evidence-grounded assistant should preserve that distinction so the analyst can assess whether a suggested step applies to the current situation.

The documented interface emphasized missing evidence as part of the output. That is an important design choice for security work, where an incomplete record can support several explanations. Making the gap visible helps the analyst choose a discriminating check and assess the remaining alternative explanations.

Retrieve operational facts through deterministic tools

Security-system integrations were exposed as deterministic tools. The assistant used these interfaces to retrieve operational facts and support its investigation. This connected the model to explicit service results with a clearer basis for validation than an inferred account of system state.

The analyst retained approval over ticket updates, escalations, and response actions. A proposed action therefore remained separate from the authorized operation. The workflow could present the relevant evidence and next step while preserving the human decision at the point where the case would change an external system.

This pattern supports accountability. Reviewers can inspect what the assistant requested, what the tool returned, and which action the analyst approved. It also creates a practical boundary for testing unauthorized requests and unexpected tool responses.

Protect sensitive institutional information

The documented implementation kept deal documents, client files, and material nonpublic information outside the model boundary. Controls included role-aware retrieval, field-level redaction, service authentication, and logging of prompts, sources, tool calls, and approvals.

These controls addressed information access throughout the workflow. A source can be relevant to a security event while still being inappropriate to expose to the assistant. Role and field-level controls allow the investigation to use permitted operational context without broadly ingesting sensitive business content.

Logging supported reconstruction of the assistant’s behavior and the analyst’s decisions. For any continuing deployment, that log itself requires appropriate access and retention controls because investigation traces can contain sensitive context. Access to those traces should follow the organization’s investigation and information-handling requirements.

Evaluate the investigation workflow

The implementation used labeled cases, prompt-injection and data-leakage tests, regression gates, quality monitoring, latency monitoring, and a rules-based fallback. These controls addressed both the assistant’s output and the behavior of the surrounding application.

Labeled cases provide examples against which the system’s summaries and suggested actions can be assessed. Adversarial tests examine whether retrieved content or user input can redirect the assistant across its authority or information boundary. A fallback gives operators an established path when the model-assisted workflow cannot meet its requirements.

The reported operational outcome was improved triage consistency and a shorter path from alert to analyst decision. The workflow brought enrichment, approved guidance, and review into the same application, reducing the fragmentation of the investigation process.

Rollout and transferable lesson

The work involved cybersecurity, application engineering, infrastructure, and operational-risk teams. Investigation runbooks were converted into reusable APIs and workflows, and the platform was rolled out in controlled phases. This connected the technical application to the teams responsible for its operating environment.

The documented result was a shipped internal investigation workspace combining alert context, deterministic enrichment, source-linked assistance, and analyst approval. It demonstrates how an agent can support a consequential workflow while its access and actions remain bounded by the surrounding system.

The transferable lesson is to make the investigation trace explicit. An analyst should be able to follow the path from alert to evidence, from evidence to a proposed next step, and from approval to the resulting action. That trace is central to both useful assistance and accountable security operations.