Nusuk Cognitive AI Assistant: a multilingual Arabic/English assistant for Saudi Arabia's Hajj & Umrah platform. The work improved digital service accessibility for pilgrims and significantly reduced operational support load.
Architected agentic workflows for Nusuk: agents with function/tool calling and multi-step reasoning (ReAct, Plan-and-Execute), and multi-agent orchestration with shared memory and human-in-the-loop control, acting on downstream APIs across multiple enterprise products.
Built production RAG on Milvus with multilingual Arabic/English embeddings, hybrid semantic search and re-ranking, grounding the agent layer in enterprise knowledge.
Delivered hybrid LLM routing between OpenAI and Arabic-specialized models for high-concurrency public-sector workloads.
Fine-tuned open LLMs (Qwen3, Llama) with LoRA/QLoRA/PEFT and served them on vLLM for low-latency, GPU-optimized inference.
Engineered end-to-end speech AI (streaming ASR, TTS and full speech-to-speech with turn-taking and barge-in) for voice-based pilgrim assistance.
Designed async FastAPI microservices with Redis caching, rate limiting, JWT and RBAC/ACL on GCP-hosted Kubernetes (Terraform/Pulumi, Helm); led AI security, including prompt-injection mitigation and tool-use guardrails aligned with government standards and the OWASP LLM Top 10.
Owned the MLOps/LLMOps lifecycle: CI/CD with canary releases, model registries, LLM evaluation and tracing, Kubeflow orchestration and drift monitoring (Prometheus, Grafana, OpenTelemetry, Evidently AI).
AI Assistant
Raseel 
