Sein Him Law

HKSoka · LLM/RAG Systems (hksoka.com/en) | Apr 2026 – Present Live production LLM platform — designed, built, deployed, and operated end-to-end. 13 models across 4 LLM providers (Claude, Grok, Kimi K3); 30M+ tokens processed; 82% of API calls fully background-automated. Built and deployed a remote MCP server (AWS Lambda) exposing HKSoka's RAG memory system via the open MCP standard — read-only retrieval, per-request audit logging; live integration with external AI clients. Multi-layer RAG memory system: user-seeded long-term context + auto-learned facts with separate retrieval quotas; nightly critical seed extraction into system prompt; bilingual embedding design for Cantonese-English retrieval; write-time quality gate + semantic deduplication at ingestion; date-aware retrieval integrates recency into ranking. Agentic model router: lightweight classifier selects model tier per message based on complexity and context; routing decisions logged per message. Token cost management: prompt caching on stable prefix — 54% cost reduction on short conversations, 80–90% on long sessions. Production billing: Stripe subscription lifecycle, token-based credit metering, referral incentive system. Stack: Claude / xAI Grok / Moonshot Kimi K3, Neon PostgreSQL, pgvector, Vercel, React/TypeScript