I'm a security and systems engineer who builds things across the whole stack โ from pre-OS firmware written in C up to multi-agent LLM pipelines โ and I care most about the parts that are easy to get wrong and hard to verify.
A lot of my work lives at the intersection of applied AI and security: orchestrating fleets of language-model agents that read code, form hypotheses, and have to prove them before anything gets reported. The thread running through almost everything I build is verification โ separate contexts for finding a thing versus grading it, reproducibility checks, calibration against ground truth, and the discipline of letting a run exit with an honest "nothing here" instead of a confident guess.
I like going low: a UEFI password lock, a from-scratch KVM hypervisor, constant-time crypto, TPM attestation. And I like shipping finished things โ CLIs, desktop GUIs, installers, CI โ not just prototypes. When I'm not doing security work I build apps and games for problems (and games) I actually care about. ๐ What I Do Multi-agent LLM systems
Design pipelines where role-specialized model agents run as isolated subprocesses, each with its own context window and rate-limit pool, coordinating through atomic file-based IPC โ no shared memory, no hidden state.
Build the thing that makes those systems trustworthy: adversarial second-pass verification, hard pass/fail gates, anchored severity rubrics, continuous calibration measurement, and drift audits.
Treat cost as a first-class concern โ per-invocation token ledgers, tiered budgets, model-tier selection, and auto-stop conditions.
See more
Systems & firmwareWrite low-level code against the UEFI firmware interface (EDK II) and the KVM API โ bootloaders that run before any OS, and a Type-1 hypervisor that loads and runs a real ELF kernel as a guest.
Hand-roll the primitives when it matters: PBKDF2-HMAC-SHA256, constant-time comparison, TPM 2.0 PCR reads for tamper detection, Secure Boot signing and MOK enrollment.
Test firmware end-to-end in automated QEMU + KVM harnesses that run on every push.
Security tooling & engineering
Build cross-platform auditors that check a machine's real state โ host hardening, VPN tunnel integrity โ and map every finding to a plain-English headline and a copy-paste fix.
Take sandboxing seriously: default-deny network bridges, egress allowlists at a proxy, dropped capabilities, read-only rootfs, and threat models written down.
Favor local-first, privacy-respecting designs where your data never leaves your machine.
Apps, full-stack & games
Ship full-stack products across web and mobile from a single codebase (React / Next.js, React Native, server-authoritative engines with time-travel-testable logic).
Build browser games and desktop utilities end-to-end, from protobuf save-format parsing to 3D model viewers to in-browser audio synthesis.
๐ก๏ธ polysec-harness โ a multi-agent LLM security-research pipeline (~45k lines of Python, 2000+ tests). Role-specialized agents rank targets, write executable proofs-of-concept, and pass through a four-gate adversarial verifier; severity scoring is calibrated against a 300-disclosure ground-truth corpus.
๐ bootgate โ a pre-OS UEFI password lock in C. It runs before any operating system loads, derives the key with PBKDF2 (100k iterations), uses TPM 2.0 PCRs for firmware-integrity tamper detection, and ships with a QEMU+KVM CI harness.
๐๏ธ wraith โ a from-scratch Type-1 hypervisor for laptop introspection and defense, built on KVM. Loads a real protected-mode ELF kernel and watches the guest from a layer below the OS.
๐ค redteam-harness โ a judge / attacker / sandboxed-target LLM red-team rig with a live D3 + Chart.js visualizer for the leaderboard, category heatmap, and cost meter.
๐ฉบ hostcheck โ a cross-platform PC security auditor: 82 read-only checks across Windows, Linux, and macOS, each tied to a fix command. leakcheck does the same for VPN leaks, with a pure-Python RFC 5389 STUN implementation.
๐พ bl2-save-editor โ a full-stack Borderlands 2 save editor: Flask API, Gearbox protobuf parsing, a Three.js in-browser model viewer, and a holographic UI.
See more
๐ mothership โ a hierarchical research swarm: a CLI that decomposes a target MECE, then spawns agents that spawn agents, each re-scoring its children's evidence before a final verifier pass.
๐ฅ group-streak โ a shared-accountability app where a whole group owns one streak. Built twice: as a Next.js + Prisma + Postgres web app, then as an Expo / React Native app for iOS, Android, and web on a Fastify + SQLite server-authoritative engine.
๐๏ธ hollow-ledger โ a monochrome browser roguelike about relics that grow across a run. Next.js / React on Cloudflare Workers with D1 + Drizzle.
๐ฌ gmail-manager โ local-first Gmail control: bulk delete, RFC 8058 unsubscribe sweeps, a CLI, and a web GUI; your token never leaves your machine.
๐ฎ rl-map-manager (Rocket League workshop-map browser), roblox-multi-instance (.NET 8 + WPF multi-account manager with DPAPI-encrypted storage), flaggle-lofi (an offline-capable daily flag game with a browser-synthesized lofi soundtrack โ play it), and keythock (system-wide mechanical-keyboard sounds synthesized per keypress).
| Area | Languages / Tools / Frameworks |
|---|---|
| Languages | |
| AI / LLM Systems | |
| Smart Contracts / Web3 | |
| Security & Low-level | |
| OS / Shell | |