Social Agent Sim — A social world model. 6 LLM agents live in a Berlin apartment building with zero personality prompts — just a 2-line bio, physics constraints, and 15 possible actions. Identity, relationships, and collective action emerge from the architecture, not from prompt engineering. The engine controls time and physics. The agents control everything else.
Claude Sonnet 4.6 · TypeScript · Model-agnostic · Environment design over prompt engineering
| Project | What it does | Stars |
|---|---|---|
| Agentic Orchestration Layer | System 2 cognitive middleware — runtime code generation replaces static data pipelines. Self-correcting triangulation protocol (SQL path vs Python path, Δ < 1%). |  |
Integration Engineer — I build real-time middleware between enterprise CRMs (Microsoft Dynamics 365) and modern frontends (Webflow, Next.js). Entity reconciliation, automated content pipelines, high-performance caching with Redis. Production systems handling live business data daily.
AI & Orchestration — Gemini, Claude, OpenRouter, Supabase (PostgreSQL + pgvector), E2B Sandboxing, Composio
Backend & Infra — Node.js, TypeScript, Vercel Serverless, Redis / Upstash, Docker, Microsoft Dynamics 365 API
Frontend — Next.js 16, Tailwind v4, Zod, SSE streaming
Payments & Auth — Stripe, Google OAuth, Supabase RLS
Philosophy — No frameworks where raw API calls give full control. No LangChain, no CrewAI. Hand-written orchestration loops. CAG over RAG — the agent writes executable code, not retrieval summaries.