Senior AI Architect / Senior LLM Engineer building production-grade agentic LLM systems, RAG platforms, AI market-intelligence pipelines, and multi-agent automation systems.

I focus on transforming real-world business and data problems into reliable AI products: routing, retrieval, context assembly, structured outputs, eval loops, observability, and production deployment. My work sits at the intersection of LLM platform engineering, multi-agent orchestration, AI safety, and market/decision intelligence systems.

• 🧠 5+ years building AI, ML, and production LLM systems
• 🏢 Senior LLM Engineer @ AIR Property, Dubai
• 🤖 Former AI Engineer, Agentic Systems @ PropTy Global
• 🏗️ Former Chief AI Officer & Multi-Agent Architect @ Novel Mind Scientist
• 🔬 Research Intern, Safe & Generative AI @ Mathis Lab, EPFL
• 🎓 B.Sc. in Computer Science, Shahid Beheshti University

📄 See my full CV

I design AI systems that are not just chatbots, but controlled intelligence pipelines with clear routing, evidence handling, structured outputs, and measurable quality.

• Agentic AI Systems
  Multi-agent architectures with task routing, memory, tool orchestration, handoff logic, fallback strategies, and production guardrails.

• RAG & Retrieval Platforms
  Retrieval pipelines with vector databases, hybrid search, reranking, context assembly, caching, evaluation, and governance.

• AI Market Intelligence Systems
  Pipelines for collecting, structuring, and reasoning over external data sources such as market data, financial signals, crypto/on-chain data, news, and domain-specific knowledge.

• LLM Evaluation & Observability
  Evals, tracing, prompt/context/output logging, quality metrics, regression tests, dashboards, and feedback loops tied to product KPIs.

• AI Safety & Security
  Red-team evaluation, prompt/tool governance, fallback policies, robust evaluation, and research-to-production security practices.

• Production RAG + agent services
• Routing-first LLM architectures instead of one universal agent
• Structured-output pipelines for reliable automation
• AI observability with tracing, logging, evals, and debugging workflows
• Market-intelligence and decision-support systems powered by LLMs
• Research-to-production transfer in AI safety and model security

• Productized reusable multi-agent RAG platforms with standardized prompts, tools, memory, evaluations, and guardrails
• Built production agentic workflows for autonomous recommendations and business decision support
• Achieved 85%+ end-to-end task completion in real-world business workflows
• Designed SLO-driven AI services with tracing, dashboards, fallback trees, and safety metrics
• Improved reliability under peak load using vector caching, retrieval tuning, and inference optimization
• Built eval/feedback loops connected to live KPIs for continuous improvement and drift monitoring
• Co-authored AI safety research accepted at ICCV 2025 and NeurIPS 2024

Aug 2025 – Apr 2026

• Architected and shipped production LLM services across RAG, agents, and decision-support workflows
• Owned model selection, prompt/agent design, retrieval strategy, evaluation harnesses, and fallback trees
• Built governed AI toolchains with tool registries, policies, approvals, and reusable architecture patterns
• Implemented reliability workflows with SLOs, error budgets, monitoring, tracing, and safety metrics
• Optimized retrieval and inference through data curation, vector caching, routing improvements, and capacity planning

Aug 2024 – Sep 2025

• Architected multi-agent systems using LangChain, custom orchestration, and RAG pipelines
• Built context-aware planning with agent-to-agent protocols, memory, dynamic routing, and escalation paths
• Productionized AI services with Docker, Kubernetes, Prometheus/Grafana, and centralized logging
• Closed the loop with automated evaluation and feedback tied to live business KPIs

Oct 2022 – Sep 2025

• Led delivery of LLM-powered agents across SaaS, healthcare, and education domains
• Designed multi-agent automation pipelines using LangChain, Celery, APIs, and knowledge systems
• Established engineering standards including design docs, ADRs, evaluation protocols, onboarding guides, and review processes
• Integrated text, vision, and knowledge-graph components into reliable AI services

May 2025 – Sep 2025

• Co-authored ICCV 2025 accepted work on data-free diffusion-based trigger inversion for Trojaned models
• Built latent-diffusion pipelines with classifier-guided feedback for exposing adversarial vulnerabilities
• Developed zero-shot and data-free defense methods with large-scale benchmarking and robust evaluation practices

Sharif University & Shahid Beheshti University | 2023 – 2025

• Prototyped secure agentic ML pipelines with RAG, routing, memory management, and evaluation loops
• Worked on AI model security, agent evaluation, and optimization research
• Contributed to publication-driven research and mentored junior engineers

Mar 2023 – Feb 2024

• Delivered a Django-based agentic recommender system with hybrid search and automated workflows
• Led Agile delivery, CI/CD, and platform iteration for production AI features

• Agentic AI
• Multi-Agent Systems
• RAG
• Routing & Context Assembly
• Structured Outputs
• Prompt Engineering
• LLM Evaluation
• Safety Guardrails
• Memory Systems
• Tool Orchestration
• LangChain
• LangGraph
• LlamaIndex

• Eval Harnesses
• Prompt/Context/Output Logging
• Tracing
• Regression Testing
• SLOs & Error Budgets
• Drift Monitoring
• Prometheus
• Grafana
• ELK
• Langfuse-style Observability

• Python
• FastAPI
• Flask
• REST / GraphQL
• Docker
• Kubernetes
• Celery
• Redis
• MLflow
• Airflow
• CI/CD

• PostgreSQL
• MongoDB
• Redis
• Pinecone
• Weaviate
• Chroma
• pgvector
• Neo4j

• AWS
• GCP
• Azure

• Python
• C++
• Java
• C#

• Routing-First Agent Systems
  Designing workflows where user intent is routed to specialized pipelines instead of relying on one generic agent for every task.

• Multi-Agent RAG Platforms
  Standardized prompts, tools, memory, retrieval, evals, fallback policies, and guardrails for reusable AI services.

• Context Assembly & Structured Outputs
  Building controlled pipelines that retrieve, rank, compress, format, and validate context before producing machine-readable outputs.

• Evaluation & Optimization
  Prompt A/B testing, routing comparisons, retrieval-quality checks, automated evals, and KPI-connected feedback loops.

• Safety & Governance
  Prompt/tool policies, red-team evaluations, fallback strategies, incident-response-ready production patterns, and robust AI security practices.

A reusable platform for building agentic AI services with routing, retrieval, memory, tool orchestration, evaluation, and guardrails.

Core ideas: LangChain/LangGraph-style orchestration, vector search, prompt governance, fallback trees, eval harnesses, and observability.

A decision-intelligence architecture for collecting and analyzing market, news, crypto, and domain-specific signals through specialized LLM pipelines.

Core ideas: external data ingestion, retrieval, scenario-specific pipelines, structured outputs, signal summarization, tracing, and evaluation.

Research and engineering work on Trojan/backdoor detection, robust evaluation, and data-free defenses for safer AI systems.

Core ideas: diffusion-based inversion, adversarial scanning, benchmark evaluation, zero-shot defenses, and production risk translation.

• DISTIL: Data-Free Inversion of Suspicious Trojan Inputs via Latent Diffusion — ICCV 2025 (accepted)
• Scanning Trojaned Models Using Out-of-Distribution Samples — NeurIPS 2024 (accepted)
• Comparison of Pre-Training and Classification Models for Early Detection of Alzheimer’s Disease Using MRI — I4C 2023

• Best Ideator — National Young Scientists Festival (2023)
• Top 0.2% National Entrance Exam — Rank 352 / 150,000 (2020)

B.Sc. in Computer Science
Shahid Beheshti University, Tehran
2021 – 2025
GPA: 3.4 / 4.0

• Persian — Native
• English — Professional

• Email: sepehrrezaee2002@gmail.com
• GitHub: github.com/SepehrRezaee
• LinkedIn: linkedin.com/in/sepehr-rezaee
• Website: sepehrrezaee.com