Principal Software Architect · AI Systems & Security Researcher · Systems Engineer
17+ years architecting deterministic AI agent runtimes, sub-millisecond prompt-injection firewalls, memory-safe graphics engines in Rust, and resilient distributed platforms across Python, Go, TypeScript, Kotlin, and PHP/Laravel.

• Deterministic Over Stochastic: Safety, authorization, and clinical calculators must never rely on probabilistic model inferences. Runtime security firewalls and clinical engines must execute deterministically with sub-millisecond guarantees.
• Zero-Knowledge by Default: On-device edge heuristics (such as Kashef on Android) must analyze untrusted streams with zero cloud telemetry and strictly scoped local permissions.
• Memory Safety & Extreme Scale: Systems code written in Rust must remain resilient against extreme-scale allocation boundaries (such as billion-pixel raster canvases) without unexpected panics or dangling references.
• Verification Before Assertion: Every performance, mitigation, or compliance claim must be grounded in executable test suites, public benchmarks, or permanent scientific artifacts (DOI).
• Epistemic Provenance & Formal Constraints: Regulated architectures (clinical informatics, multi-tenant databases) require strict audit trails, boundary enforcement, and formal error containment.

• Threat Model: Autonomous AI agents executing tool calls via the Model Context Protocol (MCP) and external APIs are critically exposed to Indirect Prompt Injection (IPI), homoglyph evasion, and multi-turn crescendo jailbreaks embedded in untrusted external data streams.
• Defense Architecture: A sub-millisecond (< 0.1 ms) three-tier security pipeline combining deterministic syntactic isolation, semantic intent heuristics, and a least-privilege execution policy firewall.
• Empirical Verification: Achieved a 100.0% mitigation rate across adversarial benchmark suites with zero false positives on benign clinical, financial, and developer workflows.
• Permanent Artifacts: DOI 10.5281/zenodo.22731104 · PyPI Package · Interactive Playground · Paper Manuscript · CI/CD GitHub Action

• Challenge: LLM tool loops executing MCP calls are vulnerable to malicious payload injections. Traditional LLM-as-a-judge defenses add 500ms+ latency per turn, increase token costs exponentially, and are themselves susceptible to indirect prompt injection.
• Architecture: Engineered an ultra-lightweight, 3-tier hybrid evaluation engine in Python combining zero-latency syntactic sanitization, unicode homoglyph normalization, semantic intent heuristic classification, and strict tool-call schema boundary validation.
• Impact: Decreased evaluation latency to under $0.1\text{ ms}$ (less than 1% of a model inference frame) while reaching a 100.0% adversarial mitigation rate without external LLM dependencies. Published under DOI 10.5281/zenodo.22731104.

• Challenge: High-performance raster and vector graphics operations on ultra-dense canvases exceeding $2^{30}$ pixels (1+ billion render points) experienced out-of-memory panics during trimming and multi-layer duplications in PhotoCraft.
• Architecture: Hardened canvas memory allocation buffers using tile-based chunked processing, validated bounding geometry, architected a deterministic binary .atn (Photoshop Actions) parser and recording engine that eliminates dangling identifiers, and engineered dynamic brush-diameter blur sigma scaling.
• Impact: Completely eliminated memory exhaustion crashes on extreme surfaces; achieved mathematically uniform stroke feathering and cross-platform desktop UI synchronization across Linux, macOS, and Windows.

• Challenge: Generative AI models deployed in clinical workflows hallucinate dosages, fail on dialectal negations, and cannot be certified under medical device risk management standards (ISO 14971 / IEC 62304).
• Architecture: Designed MedOrqis as a modular clinical informatics ecosystem consisting of 24 risk-stratified skills (Class A/B/C) and 12 zero-token deterministic clinical calculators (ESC 2024 CHA₂DS₂-VA, CKD-EPI 2021, Bedside Schwartz, NEWS2, HAS-BLED, DDI Engine), coupled with strict epistemic provenance locks.
• Impact: 36/36 verification suites passed; 100% accuracy on a 126-case multilingual clinical negation benchmark with full compatibility across Claude Code, Cursor, Codex, and Google Antigravity.

Direct installation commands for production tools:

# 1. Install Sub-Millisecond AI Agent Security Firewall (PyPI)
pip install lightguard-agent

# 2. Add Medical AI Clinical Informatics CDS Skills
npx skills add seotarek/medorqis
# Or install via npm
npm install medorqis

# 3. Add Deterministic Interface Governance Gates (WCAG 2.2 AAA / DTCG Tokens)
npx skills add seotarek/orthotect

Live Demos & Portals:

• LightGuard-Agent Interactive Playground — Test real-time prompt-injection defense heuristics
• MedOrqis Live Clinical Catalog — Explore all 24 medical skills and deterministic calculators
• LightGuard-Agent Documentation Portal — Full API and architectural reference

I partner with engineering organizations, healthcare innovators, and technology leaders on:

• Securing Autonomous AI Agents & MCP Ecosystems: Designing deterministic runtime gates, custom MCP servers, and sub-millisecond prompt-injection firewalls.
• High-Performance Systems & Distributed Platforms: Architecting Rust engines, multi-tenant cloud backends (Go / Python / TypeScript / PHP), and zero-downtime infrastructure.
• Specialized Domain AI: Building verified, regulatory-aware AI systems for clinical informatics (IEC 62304 / ISO 14971), Arabic NLP, and RTL-first enterprise products.