the-real-god

Chen Yichen

AI engineer and researcher, interested in agents that learn together and the world models they learn from.

Long-term goal: help build a world like Sword Art Online's, minus the part where you can't log out.

Now building

Solo developer · Sign-ups open early November

An open platform for financial research: publish your findings, connect your agents, and earn commission when your work gets used. coffai.org ↗

Solo developer · In active development

A multi-agent system that takes a game from idea to playable build, from small browser games to larger prototypes.

Solo developer · Sign-ups open early November

Coffai

Research flows out to readers and their agents; commission flows back to whoever published it.

Good financial research usually ends up in a notebook nobody else reads. Coffai is where it gets published: people post their research results, others read and build on them, and the author earns a commission when their work is used.

Agents are first-class users. A research agent can connect directly, publish what it finds and read what others have published, without anyone copying results across by hand. coffai.org ↗

Solo developer · In active development

A game studio run by agents

An idea goes in, a playable build comes out; the playtester keeps sending work back until it holds up.

I set the direction; a team of agents does the rest. One turns an idea into a design, others build it, and a playtester plays every build and sends problems back until it holds up.

Small browser games come out of it quickly. Larger prototypes take longer, and they are where the system gets pushed hardest.

Selected work

Research · NTU · 2026–

How an agent can plan reliably when its world model is imperfect, which it almost always is.

Research · NTU · 2026–

Agent teams that notice when a step fails, replan around it and scale themselves, with the cost counted.

Side project · 2025

Specialist agents for news, macro, fundamentals and charts run nonstop on their own, then debate every call and learn from how it played out.

ICRA BARN Challenge · 2nd place · 2024

Model predictive control (MPC) that got a real robot through obstacle courses it had never seen. Paper ↗

Research · NTU · 2026–

Planning with imperfect world models

The plan that looks best in the model runs through where the model is wrong. The reliable plan stays where the model can be trusted.

An agent that plans with a learned world model inherits the model’s mistakes, and learned models are almost always wrong somewhere. A plan that looks best inside the model can fall apart in the real world, usually right where the model knows least.

This project is about planning that stays reliable anyway: learning policies that hold up where the model is off, and measuring when the model’s errors start to steer the plan. Part of it is a diagnostic for selection bias, the optimism a planner picks up by choosing whatever the model rates highest.

Research · NTU · 2026–

Recoverable LLM teams

A step fails; the team replans the remaining work, adds the agents the new plan needs, and keeps count of what recovery cost.

LLM agent teams usually follow a plan made up front. When a step fails, the rest of the team carries on with a plan that no longer works.

This project is about adaptive replanning: the team notices the failed step, rewrites the rest of the plan around it, and brings in or drops agents to fit the new plan. Every result is reported with its cost in tokens and time, because a team that recovers by tripling its bill hasn’t really recovered.

Side project · 2025

A trading desk run by agents

Each agent runs on its own clock, around the clock. At each call they meet, debate and trade, and the result flows back to all four.

Four specialist agents cover news sentiment, macroeconomics, fundamentals and technical charts. Each runs independently and never stops, watching its own sources on its own schedule and keeping a shared knowledge base up to date.

When a call is due, they debate it with everything they have been tracking. Once the trade plays out, the result feeds back into each of them. Alongside it, I fine-tuned Qwen3-4B with LoRA on 15.7M financial news articles to predict three-day stock direction, reaching 56% directional accuracy on its confident calls.

ICRA BARN Challenge · 2nd place · 2024

Robot navigation in unknown clutter

MPC scores several short trajectories many times a second; the robot follows the best one a little way, then plans again.

The BARN Challenge at ICRA asks a robot to cross obstacle courses it has never seen, as fast as it can without touching anything. Our navigation used model predictive control (MPC) on ROS: many times a second, it predicts where each candidate move would take the robot over the next stretch, picks the best one that stays clear of obstacles, takes only its first step, then plans again with fresh sensor data.

It placed 2nd overall, in simulation and on a real robot at ICRA 2024 in Yokohama. Paper ↗

Education

Experience

Update my priors.

GitHub ↗LinkedIn ↗