Khomsan Phonsai

AI Systems Engineer · Full-Stack Developer · Optimization Specialist
[email protected] · github.com/greefeet · Khon Kaen, Thailand

AI Systems Engineer with 7+ years of production experience building high-performance, data-intensive systems. Currently architecting nectarserve — an agentic AI platform with a formal execution model designed from first principles, with auditability and safety as core architecture concerns.

Background in Structural Engineering and Computational Optimization (M.Eng, KKU) provides a rigorous, first-principles approach to systems design. Experienced with local LLM inference, vector embedding pipelines, and Claude Code as primary development tool.

Languages Rust · Python · C# / .NET · SQL · TypeScript AI / ML Local LLM inference · Vector embeddings · RAG pipelines · Anthropic API · Claude Code · LLM output validation · Hallucination mitigation AI Safety Auditable agent execution · Provenance tracking · Containment design · Formal execution modeling Data PostgreSQL · TimescaleDB · Redis · pgvector Infrastructure Docker · REST API design · Git Design & Build Webflow · Figma · Photoshop · Full-stack web (C# + .NET)
nectarserve
Status: Core engine functional — entering integration testing. Business layer in progress.
Continuum — Formal Execution Model
  • Blast Radius — three-tier containment model; bounds impact of any agent action before commit
  • Provenance — append-only audit trail of every agent action, decision, and state transition
  • WorldState — separates live facts from distilled memory
  • 10-gate pipeline — strict ordered execution: noise → thread resolve → intent → clarity → idle → mandate → constraints → blast radius → ack strategy → resolver
Practical LLM Engineering
  • LLM Output Control — enforced structured parsing so raw LLM text never reaches downstream systems unchecked
  • Hallucination Containment — mandate-gated and blast-radius-gated execution; hallucinated actions blocked before commit
  • Provider API Adaptation — model-agnostic with 6 providers via OpenAI-compatible API
Memory Architecture

Multi-tier memory system (L0 Signals → L1 Observations → L2 Knowledge → L3 Patterns) with exponential decay weighting for knowledge relevance over time. Vector search via pgvector with HNSW index, BGE-M3 1024-dim embeddings.

Full-Stack Developer
  • Built and maintained production backend systems in continuous operation over 7 years — high-volume time-series data with TimescaleDB and Redis
  • Owned full-stack delivery from schema design through containerized deployment
  • Engineered optimization algorithms from first principles to reduce data processing overhead in high-frequency pipelines
Web Developer
  • Delivered client-facing web applications with UI design and frontend–backend integration
Master of Engineering — Structural Engineering
Khon Kaen University, Thailand
Bachelor of Engineering — Civil Engineering
Khon Kaen University, Thailand

Download CV

2026 (Latest) Previous Version