FevMac

Open to new work

email github linkedin

I'm an agentic engineer.
I build with coding agents, and the tooling they run.

Fifteen-plus open-source projects. Security tooling, offensive and defensive. Privacy tooling, regulatory and technical. Developer tooling that audits what coding agents write. Full-stack in TypeScript, React, and Next.js, with Python and FastAPI. Research on AI security and safety — whether frontier model monitors can detect payloads hidden in ordinary text.

I find the gaps in modern tooling, then write the code to close them.

Approach

Research first, then research again.

I don't start building until I've exhausted what's already known. Then I brainstorm against that, go back and research whatever the brainstorm exposed, and keep the loop running until new reading stops changing my mind. Only then does architecture get decided.

From there it's a spec, a written plan, then execution — with a checkpoint between each, so a wrong turn costs a step, not a rewrite.

Then I attack it. The tools in the grid below get pointed at my own codebases: secret scanning, dependency and supply-chain audit, header and API testing. The whole loop runs under architectural decision records that make security constraints enforceable rather than aspirational. If a finding survives that, it was worth writing down.

Featured: AI security research

Covert Semantic

Can one model hide a structured secret inside ordinary text, and can another catch it? A black-box study of semantic-steganography detectability along two axes: what the monitor knows, and the form the payload arrives in. Built for the Anthropic Fellows Program application.

A defensive-measurement project. It generates a benign payload corpus and measures how monitors respond. No attack tooling.

Trained distributional detector reaches 0.94–1.00 AUC in-distribution, Holm-significant across four constructions
Held-out construction testing demoted it to a construction-family probe — reported as a negative result
Cold-prompted frontier monitors are near-blind, including on a 100%-decodable acrostic; detection tracks model lineage, not capability
LLM Evaluation Steganography Python
—
Featured: Privacy engineering

Subrosa

A privacy-engineering suite in Rust: nine Claude Code skills over a citation-backed regulatory taxonomy of 29 jurisdictional and sectoral records, bridged by a statutory dissolution map.

Compliance is a floor; selective disclosure is the ceiling.

29 jurisdictional and sectoral privacy records, each citation-backed
Statutory dissolution map bridging jurisdictional and sectoral regimes
Nine composable Claude Code skills over the taxonomy
Rust Privacy Regulatory
—
Featured: Open-source contribution

USB Army Knife — upstream contribution

Merged fix to USB Army Knife, a close-access tool for penetration testers and red teamers: ESP32-S3 boards displayed a crash dump on every first boot after flashing, when nothing had crashed.

Not my project — a contribution to someone else's.

Root cause: esptool --after hard_reset yields ESP_RST_EXT / ESP_RST_JTAG on native-USB S3 targets, which board support classified as a fault
Four lines at the root cause, with pinned-core constants verified and the RP2040 build confirmed unaffected
Proposed a future-proof whitelist refactor but kept the PR minimal, leaving the scope call to the maintainer
C++ ESP32 / Embedded Merged
View PR #274

Research & AI security

Measurement-first work on what models and infrastructure disclose.

Covert Semantic

Black-box study of LLM semantic-steganography detectability: what the monitor knows versus the form the payload arrives in.

Built for the Anthropic Fellows Program application. Defensive measurement, no attack tooling; negative results reported alongside positive ones.

llm steganography evaluation
Python
—
Credence

Exposure intelligence for the AI-infrastructure layer: leaked credentials, MCP and agent configs, git-metadata secrets, supply-chain risk.

Live verification across 16 providers and orphan-signal triage rank what to trust; SARIF dedup for CI.

llm secrets sarif
Python
—

Security tooling

Offensive and defensive tooling, ordered by phase: discovery, exploitation, detection, response.

Discovery

Dockyard

Async port scanner: CVE sigs, fingerprints, passive recon (Shodan), QUIC, Nmap pipe, script hooks — HTML / Nmap XML / Nuclei JSON.

Throughput and timing profiles for authorized scopes; audit logging and machine-readable exports.

async cve reconnaissance
Python
—
Specter

Subdomain takeover + takeover-style risks: 40+ provider fingerprints, CT/DNS/RDAP, DoH, second-order CNAME chains.

Discovery → verification before dangling assets get claimed.

dns doh enumeration
Python
—
ShadowHunter

OSINT + dark-web pipeline: Tor/clearnet ingest, Neo4j graph, STIX 2.1, NL → Cypher — attribution-first.

Hunters + crypto tracing + stealer logs on a FastAPI + Next.js stack.

neo4j osint dark-web
Python
—

Exploitation

ClaimJumper

JWT toolkit: 15+ checks, large wordlist cracking, CVE-specific chains (alg confusion, kid/JKU/X5U, null sig, JWKS spoofing).

One audit path from secret testing to forged tokens and exploit validation.

jwt cve authentication
Python
—
Stiletto

SQLi scanner: WAF-aware YAMLs, Trickest CVE sync, SARIF/JSONL for CI — GraphQL, OOB, second-order modes.

Batch URLs + CI exit codes; optional payload learning from failures.

sqli waf sarif
Python
—

Detection

Prizm

MV3 extension: storage, WebSockets, SW caches, GraphQL — 157 patterns, ML classification, live secret verification.

In-browser surface other scanners rarely touch.

extension ml analysis
JavaScript
—
Corsair

Header auditor: 60+ checks, large fingerprint set, KEV linkage, OWASP/PCI/SOC mapping, drift history, MCP remediation hints.

Letter grades + CVE tie-in; SARIF for GitHub Code Scanning.

headers compliance audit
Python
—
Restless

REST + GraphQL tester: OWASP API Top 10 (2023) checks, CVE-sourced payloads, SARIF, dashboard, Docker.

Single async pass for SSRF, BOLA, auth, GraphQL, rate limits — CI-ready.

owasp graphql api
Python
—

Response and defence

Shrike

Run live or on PCAP: eBPF flow + PID/binary tie-in, 25+ detectors — HTML (multi-tab), JSON, CLI, SARIF 2.1.0.

Timelines, IOC extraction, MITRE-mapped output for IR handoff.

pcap dfir mitre
Python
—
Vigil

Browser EVTX: Hayabusa, Chainsaw, or raw logs — 31 Sigma rules, PowerShell decode, 12 tabs, KQL/SPL/EQL/VQL export, no backend.

Logs stay local; heatmaps, lineage, and LOLBin radar for LotL-style cases.

react sigma evtx dfir
JavaScript
—
Lure

MV3 + Python CLI: additive scoring, YARA-X on .eml/.msg, intel sync — NIST/MITRE-aligned triage.

Canvas LURE for live severity across the kill chain.

mv3 phishing soc
JavaScript
—

Privacy engineering

Regulatory and technical — statutory mapping on one end, defences that run in the browser on the other.

Subrosa

Privacy-engineering suite: nine Claude Code skills over a citation-backed taxonomy of 29 jurisdictional and sectoral privacy records.

Compliance is a floor; selective disclosure is the ceiling.

privacy rust compliance
Rust
—
Kala

MV3 anti–behavioral biometrics: 22 modules (keys, pointer, motion, timing noise) — React 18 + TS; nothing leaves the browser.

Three privacy levels; sub-2ms overhead target.

react privacy biometrics
TypeScript
—

Applications & tooling

Product work and developer tooling, shipped.

Hogwash

Claude Skills suite for stripping AI defaults from writing, code, and design by two-pass manual audit rather than find-and-replace.

Find-and-replace removes the tells and leaves the structure; the audit targets the structure.

claude-skills writing audit
Python
—
North Bay

Accessible bilingual municipal site redesign. Task-oriented IA over org-chart IA; working search, focus-trapped dialogs, persistent theme, real language switching.

Unofficial concept, not affiliated with or endorsed by the City. Content is mock; behaviour is real.

accessibility i18n react
TypeScript
—
Joule

Web-first Flutter app for electricians: quotes, PDF generation, job templates, invoices, and contract workflows.

Product-first, minimal interface — built for use on site rather than at a desk.

flutter product pdf
Dart
—

Crypto & protocol

Solana and Bitcoin protocol work.

Kyma

Sound as a hardware security module: Solana seed-phrase backup inside ordinary audio, acoustic cold signing, and Solana Pay over sound.

Hackathon build (Solana Seeker, Monolith 2026) — a working proof of the idea, not a hardened wallet.

solana steganography android
Kotlin
—
Kiln

Solana → Bitcoin teleburn protocol. Burns a Solana NFT and records an on-chain proof linking it to a Bitcoin inscription. Live at kiln.hot.

~78 bytes on-chain, one atomic transaction, publicly verifiable.

solana bitcoin protocol
TypeScript
—
Capabilities

Research & AI security

LLM covert channels and detectability evaluation
Monitor robustness and evaluation harness design
Agent and MCP exposure; OWASP LLM, MITRE ATLAS

Security engineering

Detection-as-code (Sigma, SARIF, MITRE ATT&CK)
DFIR and network forensics (EVTX, PCAP, eBPF, timelines)
Offensive testing across web, API, and authentication
Threat intelligence and OSINT

Software engineering

TypeScript, React, Next.js
Python, FastAPI
Rust · Kotlin/Android · Dart/Flutter
Browser extensions (MV3), eBPF, Neo4j
Accessibility and internationalisation

ISC² CC — Certified in Cybersecurity