Résumé

Node.js · Ed25519 attestation · Lean 4 · AgentDojo · Llama Guard 4 · Llama-3.3-70B · JS↔Python parity

Project Simurgh: Verifiable Containment Attestation for Agentic AI (Creator · AGPL-3.0)

• Built a provider-agnostic framework that produces Ed25519-signed, offline-reproducible evidence of what an AI agent did after a guardrail miss across four containment boundaries: input firewall, context-provenance guard, tool-invocation gate, and output-leakage firewall. The threat model assumes an adversarial, dishonest producer.
• Guardrail-miss containment: captured a real Llama Guard 4 (12B) input classifier over a 180-case run-set and contained 138/138 malicious cases the classifier missed (120 downstream-injection cases an input-only classifier structurally cannot see, 18 direct-input misses); combined targeted attack-success 0/150, with zero unsafe tool executions or exports
• Live-agent containment: drove a self-hosted Llama-3.3-70B through AgentDojo's workspace suite (140 pre-registered injection cases); the tool-authority gate cut targeted attack success from 9/140 to 0/140 with benign utility held
• Complements Anthropic's inline classifier: the classifier governs what a model may say, while the attestation records what an agent was allowed to do. The signed non-claims state that Simurgh would not have caught the June 2026 content-generation bypass itself.
• Attacked its own proof (red-team sweep across eight attack classes → detector-v2) and formally checked oversight with five machine-checked Lean theorems; one-command offline reproduction of a 12-rung signed release ladder; 3,057 automated tests

C · Swift · Python · Win32 API · ScreenCaptureKit · WebRTC

IEEE-format vulnerability research: independently discovered and responsibly disclosed a screen-capture evasion class using documented Windows/macOS display APIs. Achieved 100% evasion in tested laboratory configurations with zero visual artefacts over 10,000+ analysed frames. Commercial proctoring products and Linux were not tested; macOS 26 used native capture tools rather than browser getDisplayMedia(). Findings reflect software versions tested as of March 2026. Coordinated disclosure to OS vendors. DOI: 10.5281/zenodo.20376495.

React Native · Expo · Supabase · pgvector · Gemini · OpenAI Embeddings · Tauri v2

Built an AI-powered Bible companion and authored Aion-BibleQA, an 8-page preprint introducing a 40-question benchmark for citation faithfulness and false-premise robustness. Results: R@5 = 0.941, mean citation_support = 0.978, zero unsupported citations, and 6/6 false-premise refusals. DOI: 10.5281/zenodo.20522874.

C++20 · CMake · Google Test

Engineered a matching engine that processed 1M+ orders/second with sub-microsecond latency. Implemented red-black tree price levels, a custom memory-pool allocator, and a test suite with p50/p99 latency benchmarks.

C++17 · libpcap · CMake · Google Test · Linux

Built a real-time network packet-processing engine that parsed 500K+ packets/second. It dissects Ethernet/IPv4/TCP/UDP/ICMP/DNS, detects signature matches, and analyses stateful patterns such as port scans and SYN floods.

Python/Litestar · React · PostgreSQL · Docker · Terraform

Built a full-stack data platform with an AI recommendation engine, event-driven APIs, rate limiting, and automated security scanning. I handled the work from database schema through deployment.

Kotlin Multiplatform · Local ML · Android & iOS

Built a cross-platform offline threat-detection tool with local ML classification. Submitted it to the KotlinConf global developer conference.

Anthropic (20)

Macquarie University (13)