Portfolio

P1. Dynamic Security Control Sets for AI-Agent Tool-Chains Using Most Restrictive Set Composition Tools can be added to agentic ecosystems without per-combination threat modeling, even under combinatorial growth by attaching policies to each tool and then compositing a control set based on execution contexts. Empirical evaluation across 120 tools.

What it enables. Production deployment of multi-tool agentic systems without exponential governance cost. Every tool invocation authorizes against the intersection of all contributing policies (catalog, tenant, engagement); adding a tool can only tighten the authorization envelope. Operators tune control strictness by context: high-assurance deployments use narrow control sets that block more chains; exploratory deployments use broader sets that admit more tools at the cost of higher residual risk. The framework provides both authorization composition and emergent-failure detection, the two properties required for safe agentic deployment at production scale.

P2. Separable Expert Architecture for Privacy-Preserving Model Personalization Per-user model personalization through device-bound LoRA adapters, expert routing bias vectors, and steering vectors. Trains personalized models without ingesting user data into the base model weights; supports GDPR compliance and data-residency requirements; preserves personalization quality through DPO-trained per-user adapters at rank 4, 16, and 20.

What it enables. Personalization at scale where the base model never ingests user data. Per-user adapters are device-bound and trained on the user's own preference signal; the base model trains on generic corpora only from a collection of LoRAs. Aligns with GDPR right-to-erasure by construction: deleting a user's proxy removes the personalization without retraining the base. Compatible with mixture-of-experts deployments and supports per-tenant or per-region adapter isolation for data-residency regimes.

P3. Verifiable Semantics for Agent-to-Agent Communication Microsoft AI & Wabash College. 2026. Link, Featured by Mustafa Suleyman, CEO, MAI.

Certification protocol that tests two agents on shared observable events and admits only terms with statistically reliable agreement into a certified core vocabulary, restricting downstream reasoning to that core. Reduces agent disagreement by 72 to 96 percent in simulation and 51 percent on fine-tuned language models; logs every certification decision to a public ledger for third-party audit; supports drift detection through recertification and vocabulary recovery through renegotiation.

What it enables. An auditable foundation for agent-to-agent communication in production environments. Every certification decision is logged in a public ledger that third parties can independently verify. Operators tune the protocol to their risk profile: safety-critical contexts use strict thresholds, yielding a smaller certified vocabulary with higher confidence in alignment; exploratory contexts use relaxed thresholds, yielding a larger vocabulary with more tolerance for residual disagreement. The framework provides both auditability and reproducibility, the two properties required for reliable multi-agent deployments at scale.