Monitoring & daily reporting
AD + firewall + SYSLOG
Custom collectors bring Active Directory and security-log evidence into a unified reporting workflow. This provides a broader operational view across identity and network activity.
Security intelligence / project 02
Custom collectors, firewall APIs and AI-assisted interpretation turn security logs into a clearer picture of what actually happened.

Collect · correlate · explain
A generic indicator description can explain why something is suspicious. Alfred asks what happened in this event: was the activity blocked, did communication continue, was there meaningful recurrence, and what can the evidence actually establish?
System design / two complementary paths
Monitoring & daily reporting
Custom collectors bring Active Directory and security-log evidence into a unified reporting workflow. This provides a broader operational view across identity and network activity.
Event-driven IOC analysis
An indicator of compromise (IOC) starts a dedicated workflow. API retrieval collects relevant firewall/SYSLOG evidence around the event, followed by classification, contextual enrichment and AI-assisted reporting.
The IOC path is separate from daily reporting. Near-real-time means analysis follows an incoming event; it is not a claim of instantaneous streaming or a guaranteed response time.
Capture the IOC trigger and its available event context.
Query current firewall/SYSLOG evidence through custom API-backed collectors.
Apply deterministic risk rules and correlate endpoint evidence where attribution is supported.
AI explains the evidence, practical significance and remaining uncertainty.
Deliver a structured incident report for human review.
Logical workflow, simplified for this case study. Collection services make the API requests; the AI reporting layer consumes their evidence. No production topology or endpoint details are shown.
From log severity to operational significance
The source alert and Alfred’s practical assessment stay distinct. A concerning indicator may describe an attempted connection that was contained, or one followed by successful traffic and stronger evidence of compromise.
Lower practical concern
A lower risk assessment still describes the limits of the available evidence. It does not prove the endpoint is clean.
Elevated practical concern
The explanation shows what changed the priority and why a human should investigate.
Higher practical concern
The assessment identifies the evidence supporting escalation and recommends human review.
Illustrative risk factors, not an exhaustive rule table. Actual classification depends on the evidence combination. Source severity is preserved; deterministic rules control Alfred’s practical tier.
The role of AI
AI turns the collected facts and computed risk into an event-specific narrative: the best-supported trigger, why it matters, what remains uncertain and what to review next.
The explanation uses the collected firewall/SYSLOG evidence and available device context. Generic web-based descriptions of an IOC are background information; they cannot establish what happened in this environment.
The language model is explanatory. It cannot override the deterministic classification or promote a confident-sounding narrative above the underlying evidence. This separates repeatable risk decisions from natural-language interpretation.
A likely trigger is not automatically proof of an internal detection signature. A hostname is not proof of a person’s identity, and a current device association is not proof of its exact location when the event occurred.
What I built
The work connects source-specific collection, event processing, deterministic evidence handling and local AI reporting into a usable security-monitoring system.
Built source-specific collection workers for directory/security monitoring and IOC intake, keeping collection logic separate from the AI prompt and reporting layer.
Used the event context from each IOC trigger for targeted firewall/SYSLOG API queries. Retrieved records become structured evidence bundles for classification and reporting.
Passed structured JSON results between workers, separating source findings, practical risk and attribution from the generated explanation.
Used a local language model to turn structured evidence and rule-based risk into a readable assessment with explicit uncertainty.
Repeated rows from one session are not automatically repeated behavior. Grouping trigger activity into episodes helps prevent a single burst of log noise from inflating practical risk.
Device correlation proceeds only when the identity evidence supports it. Ambiguous endpoint information remains unproven rather than becoming a confident identity claim.
The reporter consumes collection and classification outputs. Raw findings remain authoritative and can be checked against the explanation.
Build & validation
Investigated risk inflation caused by unrelated log activity and by treating repeated records as separate threat episodes.
Refined endpoint attribution and reporting sequence so the report can use supported device evidence while preserving uncertainty.
Used targeted replays and a fresh end-to-end event to check the corrected behavior before recording a stable release.
The result
The finished output brings together the source finding, best-supported trigger, practical risk, device context, timeline and recommended next steps in a readable PDF delivered to reviewers.
That gives the operator a reasoned account of the event, with the evidence and limitations visible.
Report structure / illustrative
What the security system reported.
The rule-based assessment and its supporting facts.
What the event means, what is uncertain and what deserves attention.
Evidence to inspect and human decisions to consider.