
Muhammad A
Agentic AI Engineer
Competenze

Consulta i miei servizi


Portfolio
Esperienza lavorativa
AI Enginner
SPS • Full time
Feb 2026 - Sep 2026 • 7 mos
• Architected a multi-agent AI compliance interviewing platform in Python/Fast API on the Gemini Live API, orchestrating up to 4 concurrent framework specialist agents (NIST CSF 2.0, SOC 2 Type II, NEPRA, VITA/HIPAA) that perform adaptive question generation, answer validation, and risk scoring over real-time WebSocket audio streaming. • Engineered the agent coordination and handoff layer - an asynchronous shared-context bus that progressively feeds every candidate answer to all active specialists, parallel inference via asyncio.gather, and a weighted result aggregator that deduplicates 150+ merged questions, resolves cross-framework conflicts, and emits a single unified risk assessment with per-framework breakdown. • Built a RAG retrieval pipeline using ChromaDB vector search over curated regulatory corpora with semantic chunking and content-hash driven re-indexing, grounding every agent finding in citable control references and materially reducing hallucinated recommendations. • Hardened the backend to enterprise security standards: AES-256-GCM encryption at rest with versioned key rotation and an ML-KEM 768 post-quantum hybrid envelope (NIST FIPS 203 / CNSA 2.0), JWT and invite-token authentication, multi-layer prompt-injection validation, rate limiting, CSP/security headers, and tamper-evident JSONL audit logging with SHA-256 PII pseudonymization - across 40+ REST endpoints and a WebSocket session gateway. • Designed and shipped an AI Application Discovery Engine - 14 independent OSINT intelligence agents fanned out concurrently into a four-stage pipeline of normalization, entity deduplication, knowledge-base classification against 500+ applications, and multi-source 0 100 confidence scoring that auto-routes high-confidence discoveries (>=70) to ingestion and ambiguous results to human review. • Built an employee discovery pipeline that maps a company's entire workforce from its domain alone - driving an asynchronous actor (run -> poll -> dataset pu