PRODUCT LEADERSHIP × MACHINE LEARNING
I ship products.
I train models.
Rasmus Rasmussen — product leader with 25+ years in games and tech, now building small language models, mixture-of-experts architectures, and the data tooling behind them.
- YEARS IN TECH & GAMES
- 25+
- YEARS IN LEADERSHIP
- 10
- MODELS PUBLISHED
- 190+
- OPEN DATASETS
- 115+
- CITED IN RESEARCH
- arXiv
01/Selected work
deep dives →Custom MoE builds
Designed and assembled mixture-of-experts models — an 8×3B (18B) and a 10B-A3B — routing across domain experts I fine-tuned myself.
LM Data Tools
A suite of data generation tools for building training and fine-tuning datasets — the factory behind 115+ published datasets.
DPO Pipeline
End-to-end Direct Preference Optimization pipeline for aligning fine-tuned LLMs, from pair generation to trained checkpoint.
ReWiz reasoning series
Reasoning fine-tunes across Llama, Qwen, Mistral and Nemo bases. One was picked up and cited in peer ML research.
AI game master
A small transformer that reads how you play and adjusts difficulty live. Built the production model at Chronicler; open tech demo in Godot.
Auto-SFT
Takes the guesswork out of LoRA fine-tuning — analyzes the dataset and derives sane training parameters automatically, from data to trained adapter.
My ReWiz-Llama-3.2-3B model was used in “Model Merging Scaling Laws in Large Language Models” (arXiv:2509.24244).
02/Track record
// the U.S. chapter — 2005 onward
-
Built an AA games startup
CHRONICLER SOFTWARE · ~3 YEARS
Spent nearly three years building the studio and owning development and training of its AI game-master model, an LLM that directs gameplay in real time.
-
Program Manager & Producer, Forza franchise
TURN 10 STUDIOS · FORZA MOTORSPORT 8 · FORZA HORIZON 5
Program management and production at Microsoft's Turn 10 Studios on two of racing's biggest titles, coordinating across studios at AAA scale.
-
Product & support leadership at global scale
VALVE
As Steam Product Manager for the Scandinavian market, I also scaled customer support from 12 people in one office to 250+ across three continents, cutting response time from two weeks to two hours.
-
Earlier: localization QA leadership & startups
RISE OF THE TOMB RAIDER · SUNSET OVERDRIVE · MORE
Loc QA lead on major AAA releases; startup roles going back to the original dot-com bubble. Shipped my own indie games on Steam — Above the Fold and Torgar's Quest — plus a couple of mobile titles.
03/Research velocity
Most of my public catalog is deliberate volume: small, cheap experiments run fast, in series, with the misses left visible. That cadence is how the serious work gets found, whether that's the MoE builds, the data tooling, or the models that end up cited. The through-line is squeezing more out of small models, for edge devices and agentic workflows, where I think the next few years of practical AI get decided. Simulations are what bridged games and AI for me, and I still build ecology sims for fun.
MODEL SERIES ON HUGGING FACE
190+ models · 115+ datasets · 12 curated collections
04/Writing
all posts →rasmus@ras.ms:~$ ./get_in_touch
Building something with AI?
Let's talk.
Product & program leadership for AI startups — from model strategy to shipped.