AI Product Engineer building production AI tooling — MCP servers, multi-model orchestration, and agent memory systems.
🟢 Open to AI Product Engineer / Full-Stack roles — AI tooling, developer platforms, agent workflows
I build agentic AI systems that go beyond chat UIs: agent runtimes, memory layers, multi-model orchestration, and MCP servers — published as npm packages developers actually run. I design systems where multiple AI models collaborate as agents: reasoning in loops, using tools, checking each other's work, and producing better outputs together.
- 🔭 Currently building TachiBot — open-source multi-model AI orchestration, live on npm
- 🧩 Exploring agent memory architectures with dokoro
- ✍️ Writing about AI architecture & multi-model systems at bypawel.com
- 🎯 3,700+ contributions in the last year
I'm building the full agentic AI stack as open source — every layer an agent needs, shipped as composable MCP pieces:
tachi-agent → agent runtime (ReAct loop, pluggable brains, tool use)
dokoro → agent memory (working / episodic / semantic / procedural)
tachibot-mcp → model council (multi-model orchestration, jury & verification)
tachibot.com → docs (guides & reference for the whole stack)
🤖 tachibot-mcp — the council: multi-model orchestration over MCP
Problem: LLM outputs are often accepted without independent verification — hallucinations slip through. What I built: A TypeScript MCP server that runs GPT, Gemini, Grok, Qwen, Kimi & more in parallel so they critique and cross-check each other's work before you act on it. 50+ tools, multi-model jury synthesis, configurable profiles.
npm install -g tachibot-mcp🧠 dokoro — the memory: multi-layer memory for AI agents
Problem: Most agent "memory" is a toy key-value store that forgets everything between sessions. What I built: An MCP memory server with working, episodic, semantic, procedural & affective memory layers — inspectable, persistent memory for Claude Code and any MCP client.
⚡ tachi-agent — the runtime: local-first agentic loop
Problem: Agent workflows usually depend on hosted APIs and hidden state. What I built: A local-first ReAct agent that fuses dokoro memory + tachibot multi-model council over MCP. Pluggable brains: Qwen via Ollama, Nous Hermes via vLLM/llama.cpp, or any OpenAI-compatible endpoint — runs 100% local.
🌐 tachibot.com — the docs: guides & reference for the stack
Documentation hub for the Tachi Stack — setup guides, tool reference, and workflows for orchestrating GPT, Gemini, Perplexity, Grok, Qwen, Kimi & MiniMax.
AI Engineering — agentic AI (ReAct agents, tool use, multi-agent workflows), MCP servers & clients, multi-model orchestration, agent frameworks (LangGraph, Mastra, CrewAI), agent memory systems, LLM observability & evals (Langfuse, Evalite), local-first LLM workflows (Ollama/Qwen), OpenAI / Anthropic / Gemini / OpenRouter APIs
Full Stack — TypeScript, React, Next.js, Node.js, Python, Django, GraphQL & REST API design, state management (Redux, Zustand), data viz with D3.js, E2E testing with Playwright, open-source package maintenance
I write about AI architecture, multi-model orchestration, and building AI-powered applications.
Open to AI Product Engineer and Full-Stack Engineer roles — especially teams building AI tools, developer platforms, agents, or production LLM workflows.