profile
The operator behind the tools
method
Three rules, learned the expensive way
Every project in this portfolio is held to these. Each exists because it was violated once, and each is enforced by a command — not a slogan.
selected work
Four exhibits, with their refutations attached
What each project claims, the command that reproduces the claim, and — stated by the author — exactly what it does not do.
coverage
OWASP Agentic AI Top 10 · 2026
covered by at least one attack in the library blind spot — no attack yet, printed in every report
verification lab
Run the algorithms, right here
A faithful JavaScript reimplementation of the documented detector logic — same signals, same thresholds, same guards — labelled as such. It is not the shipped Python artefact.
The calibrated two-layer detector from secret-sweep: pattern rules first, then an entropy layer with three structural guards. Evidence is masked in the output — this page never prints a full token.
The control plane of zero-trust-ai-gateway: RBAC resolves the role, signature inspection reads the payload, and a refusal is explicit — a 406 with the matched signature — never a silent drop.
Signature table
Gateway decision
Every number below is produced by
python3 scripts/calibrate.py against —
files, lines — and re-produced in CI on every commit.
610 → 23
lines flagged on the corpus by the naive entropy > 4.5 test versus the calibrated detector — 26× less noise, and the naive test still missed real credentials.
0.0076%
false-positive rate (23 of 304,467 lines). A detector that cries wolf gets switched off within a week; this one is CI-usable.
< 24 chars
the structural blind spot that started the project: entropy > 4.5 cannot fire below 24 characters, so a 20-character AWS key ID was undetectable by construction.
Measured recall — planted-secret corpora
Corpora are generated, not harvested: 99.5% honestly means “99.5% of the shapes modelled”. Stated in the ledger, §09.
credentials
Competencies, tagged by evidence
Every line says what stands behind it: professional practice, a public repository, or formal certification. No unmarked claims.
Certifications ·
Education
experience
Three roles, exact dates, names withheld
supporting work
Control plane & backend systems
Smaller scope than §03, same standard: each with its stated limitation.
known limitations
What this work does not do
Stated openly, because a portfolio that hides its gaps is marketing. Every row is a fact a reviewer would otherwise have to discover.
path
The order things were built in
From CS50 pointers to the adversarial counterpart of everything before it.
contact
Let’s talk security
Open to security engineering, detection engineering and AI-security roles — and to collaborators who hold their tools to a measurable standard. References, engagement letters and full employer details on request.