Inspiration
Security and identity teams deal with high volumes of access logs where risky patterns—such as unexpected privilege escalations, sensitive data exports, repeated failed logins, or anomalous source IPs—can easily get buried in routine activity. At the same time, relying solely on external AI APIs without fallback handling creates friction when cloud services experience high traffic or when teams want a structured, repeatable way to triage logs. The broader vision behind sentinelGuard also stems from privacy-conscious security defense: keeping sensitive data and decision logic controllable locally while pairing structured analysis with resilient workflows.
What it does
sentinelGuard is an AI-driven Enterprise IAM compliance and access log triage assistant built with Streamlit. Users can:
- Upload access log files in CSV format and preview the dataset directly in the browser.
- Enter their Gemini API key at runtime (never hardcoded in the codebase) and run an automated compliance audit.
- Receive a structured audit report covering an Executive Summary, Detected Anomalies / Security Risks, and Recommended Remediation Actions.
- Automatically retry transient
503high-demand errors and fall back to a deterministic local rule-based audit engine if503errors persist. - Download the generated compliance report as a
.txtfile (sentinelguard_audit_report.txt).
How we built it
We built sentinelGuard as a clean, lightweight Python application using Streamlit, pandas, and the google-genai SDK (gemini-3.8-flash):
- Interactive UI & CSV Ingestion: Streamlit handles file uploads, dataframe previews, runtime password-masked API key entry, progress spinners, and one-click
.txtreport downloads. - Resilient AI Auditing: Log data is formatted for analysis and sent to Gemini with up to three attempts using short backoff delays when
503responses occur. - Deterministic Local Fallback: If
503errors continue,generate_local_fallback_reportinspects log entries locally for indicators such as successful privilege escalations, successful data exports, failed access attempts, and non-private IP prefixes so an audit report can still be produced.
Challenges we ran into
- Handling transient API load (
503errors): Live model endpoints can experience temporary high demand, which could interrupt an audit or demo if there is no retry and fallback path. - Keeping credentials safe and setup simple: Avoiding hardcoded API keys while keeping the app straightforward to run locally via
python -m streamlit run app.pyrequired careful state and input handling. - Balancing AI depth with deterministic checks: Designing the fallback rules so they flag clear high-signal events without overpromising full enterprise SIEM/IAM correlation required keeping the scope focused and transparent.
Accomplishments that we're proud of
- Delivering an end-to-end, working prototype with CSV preview, runtime API key handling, structured report generation, and
.txtexport. - Implementing automatic retry backoff plus a working local rule-based fallback so the tool remains usable even during
503high-demand spikes. - Publishing a clean, public repository (
sentinelGuard) withapp.py,requirements.txt, andREADME.mdand zero hardcoded secrets.
What we learned
- Reliability and graceful degradation matter just as much as model output—especially during live evaluations or high-traffic moments when a local fallback keeps a workflow moving.
- Clear, scoped prototypes with honest boundaries are much stronger than overcomplicated designs that break under time pressure.
- Practical IAM and access-log triage benefits from combining human-readable summaries with concrete remediation checks (such as verifying privilege approvals, reviewing sensitive exports, and monitoring failed logins on sensitive resources).
What's next for sentinelGuard
- Schema normalization & larger log support: Normalize common column name variations automatically and add chunking/filtering for larger log files.
- Richer detection rules & IAM integrations: Expand local checks (such as repeated failures