I design, code and ship AI agents that have to work in production: agent memory, RAG that cites its sources, tool calling, guardrails, and the evals that decide when a change is good enough to ship.

• Lead the architecture of a production AI agent platform at Alien Intellect: migrated the agent from five separate contexts to one, shipped persistent per-user memory and a RAG pipeline with real citations, built the guardrails that keep the agent out of its own prompts and infrastructure.
• Own the bar for done. Prompt and context engineering, an eval suite (LLM-as-judge plus regression cases) that gates every agent change, and a Definition of Done that means "running on production with eval evidence".
• Grew the engineering team from 2 to 11 in a year. Wrote the hiring rubric and two-stage interview flow, run onboarding and coaching for every joiner, and built the delivery habits (sprint cadence, DoD, QA flow) that keep a young product shipping.
• Five years of backend at scale before that, at Cafe Bazaar: a 10,000+ RPS service on the critical path for 51M+ users, a Python-to-Go migration across 15+ services with zero downtime, and early LLM adoption that cut review-response time 192x.

I start from the why and the what, then build the how. Stack today: Python, Go, LangGraph, Claude Agent SDK, Kubernetes on AWS.

AI systems

Languages and frameworks

Data and messaging

Infrastructure

• From Five Contexts to One: What Unifying a Production AI Agent Taught Me
• RAG That Cites Its Sources
• Agent Memory That Survives the Session
• Challenges in Developing Bazaar's Update Service · how a 10,000+ RPS critical-path service is built and kept fast
• Ten Years of AI at Cafe Bazaar: From Text Mining to LLMs

More on the blog.

• EdgeLinker: Practical Blockchain-Based Framework for Healthcare Fog Applications · IEEE CPSAT 2024, first author · 10.1109/CPSAT64082.2024.10745419 · arXiv:2408.15838
• Blockchain-Based IoT Framework with In-Depth Security Analysis and Performance Benchmarks for Real-World Healthcare Fog Applications · Scientia Iranica, Transactions D, 2025 · 10.24200/sci.2025.66018.9798
• Energy-Aware Workflow Scheduling in Fog Computing Using a Hybrid Chaotic Algorithm · The Journal of Supercomputing (Springer), 2023 · 10.1007/s11227-023-05330-z

• pipeline-mips-verilog · a five-stage pipelined 32-bit MIPS processor in Verilog, every hazard resolved with stalls
• CPP-object_oriented-tutorial-with-project · the C++ object-oriented course I wrote for the high-school students I taught
• userProfile · a gRPC user-profile service in Go built for heavy load, the patterns behind Cafe Bazaar's production services
• Pong-Client-Server-Monitor · an online Pong game in Python with a real-time match monitor

Happy to compare notes on production AI agents, system design, or teaching. The door is open.