AI & operations · 2026
SignalDesk AI
Turn noisy alerts into clear next steps. Keep the data local.
My contribution
I built the redaction, classification, explanation, routing, retrieval, container and test path.
Evidence
60-incident controlled corpus, 81.7% macro-F1 cross-validation result, repository and 61-second walkthrough.
Scope & limitations
The dataset is intentionally small; predictions are advisory and require human confirmation.
- Python
- NLP
- scikit-learn
- Docker
- GitHub Actions
- Trivy
See it in action.
01:01 · Visual walkthroughAbout this demonstration
A concise visual walkthrough built from the repository's verified privacy boundary, six-class model, explanation features, runbook routing and similar-case retrieval. On-screen subtitles explain every step, accompanied by a soft motivational beat.
Receive alert -> redact private data -> classify locally -> explain evidence -> route team and runbook -> retrieve similar cases -> confirm as a human.
The problem & approach
The problem
Operational alerts arrive as messy text and may contain data that should not be copied into an external AI service. I wanted a local system that could accelerate the first triage decision while showing responders why it suggested a route and where human confirmation is still required.
The approach
SignalDesk converts an unstructured incident alert into an explainable category, responder team, severity, runbook and set of similar resolved cases. It removes emails, IP addresses and common secret formats before inference, processes requests in memory, and keeps the complete decision path local rather than sending operational text to an external model.
How it works
- 01Incident text
- 02PII and secret redaction
- 03TF-IDF classifier
- 04Feature-level explanation
- 05Team + runbook routing
- 06Similar-case retrieval
- 07Human confirmation
- I trained a six-class NLP model across authentication, compute, database, deployment, network and security incidents.
- I added deterministic redaction for emails, IP addresses and common secret formats before text reaches vectorization.
- I exposed positively weighted input features so the routing decision can be inspected instead of presented as a black box.
- I connected every category to a responder team, severity and runbook while retaining human confirmation before operational action.
- I implemented similar-case retrieval so responders can compare an alert with historical incidents and their resolutions.
- I added health, model-metadata and feedback endpoints around the responsive incident workspace.
- I containerized the service as a non-root, health-checked workload and added CI tests, image builds and blocking Trivy repository scans.
Technical decisions & lessons
Local inference
I kept incident text on the host and avoided an external model boundary for operational data.
Deterministic redaction
I removed known sensitive patterns before any model feature is created.
TF-IDF + logistic regression
I chose a transparent baseline whose weighted features can support an explanation.
Similar-case retrieval
I paired classification with concrete historical context a responder can evaluate.
Human confirmation
I kept the model advisory because the bundled corpus is intentionally small and unfamiliar incidents need caution.
Challenges
- The bundled 60-incident corpus is intentionally small, so I report cross-validation results without presenting them as production-scale accuracy.
- Redaction must happen before vectorization and retrieval; applying it only to the displayed output would leave the sensitive processing boundary unchanged.
- The interface needed to show evidence and uncertainty clearly enough that a responder can disagree with the route.
What I learned
- I learned that privacy architecture is about data flow order, not just masking text at the end.
- I learned how interpretable features and similar cases can make a small classical model more operationally useful.
- I learned to report evaluation and dataset limits beside the metric instead of hiding them behind a confident interface.
- I learned how model checks, container hardening and dependency scanning fit into one delivery path.
Privacy-first triage path · system flow
alert -> redact PII + secrets
-> classify + explain
-> route team + runbook
-> retrieve similar cases
-> human confirmation