Bringing personal super intelligence to every worker in the world.
CTO based in Berlin. I build things at the intersection of AI and product — autonomous agents that manage engineering teams, voice bots that handle real phone calls, developer tools that make Claude Code actually useful in production.
I ship fast, open source a lot, and believe the best software is built by small teams with high leverage.
Here is some of my work on hard problems in agentic AI — memory, intent, tool selection, and autonomous engineering. These come from building production agents, not from theory.
| Paper | Topic | Summary |
|---|---|---|
| Engram (v3, Mar 2026) | Agentic Memory Pathology | A pathology-first framework for diagnosing memory failure in AI agents. Defines an amnesia taxonomy (temporal, source, interference, encoding, retrieval, consolidation, prospective) validated against three production systems: Maia, hive, and Deepak. Introduces the RAG-Amnesia Scale and EngramRecord encoding. |
| iCPG (v8, Mar 2026) | Intent-Augmented Code Property Graph | Reframes a class of coding agent "hallucinations" as specification drift — measurable divergence from intent. Proposes ReasonNodes with formal contracts (preconditions, postconditions, invariants) and 6-dimension drift detection. I did it primarily to make claude code work better. |
| Mnemos (v1, Apr 2026) | Task-Scoped Agent Memory | A framework for how agents acquire, organize, compress, and hand off knowledge during a single task. Addresses context wall crashes in long-running Claude Code sessions with typed MnemoNodes, a 4-dimension fatigue model, tiered REM consolidation, and SkillNode promotion for reusable patterns. |
| Lexon (v1, Apr 2026) | Semantic Tool Binding | Solves tool selection accuracy collapse at scale. A two-tier routing pipeline with multilingual embeddings, structured disambiguation, and a personalization layer that learns user vocabulary over time. Integrates with Mnemos, iCPG, and Engram to form a complete agentic cognitive stack. |
| Telos (v1.1, May 2026) | Intent-Grounded Testing | Reframes testing from "does the output match the spec?" to "does the artifact serve the intent?" Models the lossy chain from real intent to behavior, defines 8 intent-failure modes (IF-1 through IF-8), and runs three test planes autonomously — including whether the spec itself is wrong. Cross-references iCPG (intent governance), Engram (cross-session memory), and Polyphony (decomposition closure). Operational layer: DAE (Dynamic Autonomous Evaluation). |
| DAE (v1, May 2026) | Dynamic Autonomous Evaluation | Operational evaluation substrate for autonomous agents. Continuous 5-stage pipeline (Capture→Score→Compare→Gate→Learn) with dimensional rubric scoring, regression baselines, risk-tiered adaptive sampling, and multi-judge panels. Agent-evaluator isolation as a structural design constraint. Operational layer for Telos's intent-grounded testing. |
| Maggy | Autonomous AI Engineering Agent | Released at claude-bootstrap. A local-first, self-improving engineering agent with multi-model orchestration (Claude, GPT-5, Gemini, Kimi, DeepSeek, Qwen), 5-level closed-loop control, process intelligence from CI/PR/deploy signals, and Maggy Mesh — a P2P network for sharing team learning across developer instances. |
These papers form a coherent Agent Architecture Series: iCPG governs intent in code, Mnemos governs task memory, Engram governs cross-session memory, Lexon governs tool resolution, Telos governs intent-grounded autonomous testing with DAE as its operational evaluation layer, and Maggy orchestrates all of it into an autonomous engineering platform.
Autonomous AI Systems — Most of my recent work is about making AI agents that actually do real work, not demos. Zoro is an autonomous engineering manager that runs as an iTerm2 extension — it monitors tickets, routes work to Claude Code sessions, detects error loops, and runs a web cockpit for oversight. claude-bootstrap (529+ stars) is the opinionated project scaffold I use to make Claude Code reliable across all my projects.
AI-Native Developer Tools — Hive is a standalone AI command center for SaaS — it manages budgets, creates tasks, makes strategic decisions, and coordinates between AI agents and humans. Halo brings Claude Code to the desktop. voxy is a voice-controlled terminal assistant.
Voice & Conversational AI — AIVoiceBot is a complete voice bot service handling inbound and outbound calls. realtime-transcription does live audio-to-text across languages. voiceover generates AI narrations for screen recordings.
MCP & Integrations — mcp-linkedin-server (50+ stars) is an MCP server for LinkedIn automation. I've built crawlers, search engines, proposal generators, and various connectors between AI and the tools people already use.
Alongside open source, I have worked on...
Enterprise CX Platform — a customer experience platform processing millions of survey responses. I managed/contributed to the full stack: backend services, frontend apps, Shopify integrations, and a migration from legacy to a modern multi-tenant architecture. Multi-provider integrations (Salesforce, HubSpot, Intercom). Team of engineers across backend, frontend, and DevOps.
End-to-End AI Marketing Agents — an AI-native marketing platform built around a fully autonomous AI agent that controls the entire system. The agent runs campaigns end-to-end: strategy, brief co-creation, content generation, creative production, and analytics — all via streaming chat, autopilot mode, or proactive nudges. It conducts interviews and outreach over email, WhatsApp, and live voice meetings with real-time turn-taking and TTS. Multi-agent architecture with hired sub-agents per brand, automated email cadences, MCP integrations (HubSpot, Salesforce, Google Ads, social platforms), and a full copilot toolkit with tools, skills, and context-aware planning.
AI-Powered Learning & Transformation Platform — an ed-tech platform with a suite of AI products: synthetic interviews, knowledge sprints, AI-generated podcasts, micro-learning, chat-based tutoring, content generation, onboarding companions, and a full transformation suite for organizational change. Simulation agents, ITSM process automation, and a unified platform serving it all.
I wrote a No-Agile Agile Manifesto and an Organization Consciousness Protocol because I think most process is theater. What actually works:
- Small teams, high autonomy. One engineer with good tools beats a squad with a Jira board.
- Multi-model, not single-model. I don't use one AI — I built a 9-tier routing system that classifies every task and delegates to the cheapest capable model. Qwen3 for lookups (
$0), DeepSeek Pro for implementation ($0.44/M), Kimi for review, Gemini for multimodal and deep research, Codex for bulk generation, Claude for architecture and security. DeepSeek handles ~80% of coding; Claude is reserved for what actually needs judgment. This isn't cost-cutting — it's about using the right tool for the job. Why pay $15/M for a typo fix? - Memory is the moat. Every AI coding tool loses context on compaction. Mnemos doesn't compress blindly — it tracks why each memory node exists with typed eviction policies, measures fatigue across 4 dimensions, and writes checkpoints before things go wrong. Codex and Claude Code fire-and-forget; Mnemos preserves intent.
- Autonomous, not assisted. Maggy doesn't wait for me to ask. It auto-discovers untested code and generates test suites. Background heartbeats scan competitors and refresh the task inbox. After significant changes, a Stop hook asks qwen3 whether a multi-model review (DeepSeek + Kimi + Codex in parallel) is warranted — and triggers it autonomously. The agent decides when it needs help, not me.
- Ship first, abstract later. Three similar lines of code is better than a premature abstraction.
- Test intent, not just output. Every testing framework today asks "does the output match the spec?" Telos asks a harder question: "does the artifact serve the intent?" A green test suite doesn't mean the spec was right — it means you built the wrong thing correctly. Telos tests the spec itself, scores how much intent survives each translation link, and detects when the tests themselves have been captured by a proxy. Autonomous testing that tests the test.
Python TypeScript FastAPI Claude API Agent SDK MCP SQLite PostgreSQL React Next.js Shopify iTerm2 API WebSocket GraphQL