AI Engineer

Interactive entertainment media — and how to build it responsibly.

My career is a single thread: how do interactive digital services engage people without exploiting them? I spent a decade designing engagement and retention systems for live service games with millions of users, navigating the tension between what drives KPIs and what's actually healthy for the people using the product. Now I'm bringing that same thinking to conversational AI — building systems that engage through genuine depth, not manipulation.

Currently building a voice-activated Star Citizen AI copilot as a capstone for Masterschool's AI Engineering program — LangGraph multi-agent architecture, live game data APIs, RAG knowledge base, voice I/O pipeline, and full observability via LangSmith.

Lübeck, Germany · Canadian · German work permit

The through-line

Every company I've worked for optimizes for engagement, retention, and monetization. At Supercell I managed products with millions of daily users. At PopReach I increased revenue 50% in four months. At Hothead I built the LiveOps infrastructure from scratch. The metrics were always the same: DAU, retention curves, ARPDAU, LTV.

The harder question was always the ethical one. When is an event tuned so aggressively it's borderline abusive? When does a monetization mechanic cross from fair value into exploitation? How do you hit revenue targets without designing systems that prey on the people who trust your product? I spent years navigating that tension — pushing for engagement through genuine depth instead of predatory mechanics. It wasn't always a popular position.

Conversational AI has the same tension, except sharper. These systems build relationships with real people — and attachment isn't a failure mode here, it's the point. The harder question is what that attachment is for, and what the system does when it notices dependency forming instead of just rewarding it. I'm building AI systems that engage through memory, earned trust, and real utility — a secure base, not a hook. Same fight, different medium.

2025–2026
AI Engineering
Masterschool · Capstone Project
Career pivot into AI engineering. Building a voice-activated Star Citizen AI copilot — LangGraph multi-agent orchestration, RAG, live API tools, voice I/O, OCR, LangSmith observability. Python, FastAPI, LangChain, LangGraph.
2022–2023
Game Lead — Live Operations & Monetization
Supercell · Helsinki
Game Lead for two products with millions of daily active users. Managed a multi-discipline development team. Created turnaround strategy for underperforming products. Roadmap ownership, P&L management with FP&A, game economy balancing. Supported M&A team analyzing potential acquisitions and studio business strategy.
2020–2024
Senior Product Manager
PopReach · Lübeck & Remote
Led multiple mobile products across international teams. Increased revenue 50% in 4 months through engagement strategy and audience segmentation. Developed roadmaps and product strategies, trained teams on LiveOps and analytics best practices.
2016–2019
Live Operations Producer
Hothead Games · Vancouver
Built LiveOps from scratch — JIRA workflows, production pipelines with cost evaluations, P&L reporting, community management via Discord and forums. Established KPI frameworks and forecasting for executive reporting.
2006–2015
Game Developer
Goodgame Studios · Dramaforum · Ganz · Hard Circle
C#, Unity, game systems. Started as a Flash developer, progressed to Senior Unity Developer at Goodgame Studios and Programming Team Lead at Dramaforum. Learned how systems shape user behavior from the code side.
From Games → AI
Engagement Without Exploitation
Progression systems, retention loops, economy design — built with the understanding that the metrics that make shareholders happy and the patterns that are healthy for users are often in conflict.
From Games → AI
P&L and Unit Economics
Revenue modeling, cost-per-user analysis, monetization design. Knowing what a feature costs to run, when it pays for itself, and how to be profitable without being predatory.
From Games → AI
User Psychology at Scale
Behavioral analytics, segmentation, KPI frameworks across millions of users. Understanding not just what users do, but why — and when "why" means the system is working for them vs against them.
From Games → AI
Live Operations Discipline
Products that run 24/7, serve millions, and evolve weekly. Monitoring, incident response, and shipping under the constraint that real people are depending on you right now.

Star Citizen AI Copilot

A voice-activated gaming AI assistant built on LangGraph multi-agent architecture. The user talks while playing — the copilot responds with live game data, expert knowledge, and advice personalized to their ship and playstyle. JARVIS for Star Citizen.

Building June–September 2026 Data infrastructure from prior project API integrations confirmed + keys obtained
Orchestration
LangGraph Multi-Agent
Four specialized agents — RAG, Tools, Memory, and an orchestrator that routes queries via two-tier classification (deterministic for structured input, LLM router for natural language). Parallel agent execution. Full trace visibility in LangSmith on every call.
Voice Pipeline
Push-to-Talk → Voice Response
Local speech-to-text via faster-whisper on CPU (no GPU contention with the game). Server-side processing. ElevenLabs TTS with streaming — playback starts before full audio is generated. Sub-4-second voice round-trip target.
Live Data
Real-Time Game APIs
UEX Corp API for live commodity prices and trade routes. CStone for item locations. Local PostgreSQL with game item database synced from SC Wiki. Multi-step tool interactions with cached results for follow-up refinement.
Context Awareness
Persistent User Memory
User memory tracks ships, playstyle, preferred locations, and economic philosophy across sessions. A Prospector pilot and a Reclaimer crew get fundamentally different advice from the same question.
Novel
RS Signature Lookup
Star Citizen's scanning assigns deterministic signature numbers to entities. Tesseract OCR proved too fragile on the HUD font — replaced with a local VLM for accuracy while a custom CNN trains as the production target: small, fast, safe to run client-side without competing with the game for resources.
Novel
Expert Knowledge RAG
Wiki-sourced ship data plus curated expert gameplay knowledge written from real experience. The competitive moat — community data is free, expert meta-knowledge about actual gameplay strategies is unique.
LangGraph LangChain LangSmith FastAPI Python 3.11+ PostgreSQL ChromaDB Claude Haiku Ollama faster-whisper ElevenLabs VLM (local) CNN (training) nomic-embed-text-v2 SQLite
4
Agents
<4s
Voice Round-Trip
<0.5s
RS Signature Lookup
3
Live Data APIs

Thin Python client runs alongside Star Citizen — push-to-talk voice capture, OCR for ship detection and RS signatures, all on CPU to avoid GPU contention. Server on Hetzner handles agent orchestration, RAG retrieval, API calls, and TTS synthesis. LangSmith traces every agent decision for full observability.

Built as a solo developer in three months, using LangChain/LangGraph as the orchestration framework. Designed to work as a real product with paying users — not just a portfolio demo. Personality DLC and tiered subscriptions provide the monetization path.

Writing

Technical deep-dives into the problems I'm solving and the decisions I'm making.

Let's talk

I'm looking for AI engineering roles where engagement design, ethical product thinking, and technical depth all matter. Particularly interested in conversational AI, agentic systems, and teams that care about building things responsibly. Open to full-time, contract, or compelling collaborations.