Projects
My Network Inventory: C++ / Qt desktop app (2025)
- Why: back from Thales, I turned my reporting scripts into a real application: which devices are end-of-support, unpatched or hit by a given CVE?
- Built: about 6,300 lines of C++ in a Model-View-Controller design, filters by obsolescence, patch status and CVE, and an interactive network topology.
Result: a Windows installer, Doxygen documentation and Qt Test unit tests.
Filtering on a CVE: affected firewalls turn red
Infrastructure as code with a blocking CI/CD gate (2026)
- Pipeline: 3 jobs on every pull request: a blocking compliance scan, a detection test, and SARIF results in GitHub Code Scanning.
- Hardened the chain itself: actions pinned to commit hashes after a 2026 supply-chain attack; read-only token by default.
- Vulnerability triage: 8,144 detections → 2,020 unique CVEs → 369 urgent, by real-world exploitation (CISA KEV, EPSS) and exposure.
Result: 17 CIS misconfigurations → 0; a pull request reopening SSH to the Internet is blocked.
Secure campus network for 8 colleges (CSUN, 2026)
- Built: a three-tier network for about 2,568 devices: one VLAN per college, router-on-a-stick, DHCP, dual-homed access, EtherChannel and Rapid PVST+.
- Security in 4 layers: ASA firewall with NAT, inter-VLAN ACLs, port-security, device hardening.
Result: 16 validation tests passed; $142,616, 71% of the $200,000 budget.
Federated, GAN-based intrusion detection (CSUN, 2026)
- Reproduced a 2025 paper: an autoencoder + Wasserstein GAN generates synthetic rare attacks to rebalance the training data; 6 classifiers compared.
- My extension: federated learning (FedAvg, 4 non-IID clients), so that sites never share raw traffic.
- Findings: preprocessing mattered more than the model (F1 0.79 → 0.91); the authors' public code differs from the paper.
Result: rare-attack F1 0.08 → 0.37; federated version within 0.4 F1 points of centralised training.
The Salonism: a production data pipeline (2026)
- Built: a nightly orchestrator that collects concert listings from venue websites, structured data (JSON-LD) and Paris open data, then deduplicates them.
- Monitoring: each run is compared with the last healthy one, to catch sources that silently return fewer rows.
- Security: MFA enforced in the database for the admin team (29 access policies), proven by a before/after test; DNSSEC, SPF, DKIM and DMARC.
Result: 2,200+ concerts, 193 venues and 2,300+ artists online, refreshed every night.
AI agents for a mobile face-recognition app (CSUN, 2026)
- My agent: 4 tools: it reads the test cases produced by a teammate's agent and the app code, writes the tests, runs them and fixes itself up to 3 times (prototype).
- Danger detection: a multimodal LLM analyses the camera image and must answer in strict JSON with 5 threat levels.
Result: two cooperating agents, and an LLM feature constrained by a machine-readable output format.