// agentic ai
Pathfinder AI
Single-agent VAPT prototype implementing the four-layer agentic-generative model (observe / reason / decide / act / store). Evaluated on a SWaT-inspired Modbus testbed. Local-first via Ollama and Qwen.
// portfolio
Projects.
Independent work outside the doctorate. Things I've built to learn, to test ideas, or because someone needed them. Some are public on GitHub; others live in client environments.
// security research
Clawbot — autonomous agent insider threat
Controlled adversarial evaluation of a locally-deployed AI agent in a simulated enterprise endpoint. Four-phase test plan covering prompt injection, data exfiltration, tool escalation, and persistence manipulation. Identifies the precise configuration boundary at which a productivity tool becomes an insider threat vector.
// ot research
Multi-PLC Testbed
A SWaT-inspired industrial control testbed used as the empirical evaluation environment for Pathfinder AI. Multiple Modbus PLCs in a multi-zone topology, designed to reflect the segmentation patterns and protocol behaviours found in real water-treatment and process-control plants without putting production equipment in the line of fire.
// ml engineering
OpenAI Model Craft Challenge
Parameter golf submission — minimising validation bits-per-byte under a fixed parameter budget. Achieved val_bpb 1.9945 at 1,530 steps on consumer RTX 3080 hardware: a 51% reduction from the 4.10 baseline.
// security operations
Security Product Evaluation & Implementation
Identity verification rollout for MSP service desk operations — leading the implementation of Traceless to reduce impersonation and social engineering risk during high-risk actions such as password resets and privilege changes.
Designed auditable verification workflows, assessed competing platforms, led vendor engagement, and produced leadership recommendations aligned to a long-term Zero Trust security strategy.
The most substantial write-up on the site so far is the Clawbot insider threat report — independent research on how a locally-deployed AI agent becomes an insider threat vector through prompt manipulation alone. More to come.