Cornucopia
An open-source Chinese financial LLM series and a lightweight vertical-domain training stack.
Explore the flagshipAI Full-Stack Engineer · Norway
Cornucopia.AI
Building open-source LLM & agent systems — from post-training to production.
I work across the LLM stack: post-training, agents, retrieval, and the engineering that puts them into production. I care about systems that hold up outside the demo.
- Stars earned
- 736
- Public repos
- 25
- Followers
- 26
金鼎
The Cornucopia emblem
01 — About
Build for the real world,
not the demo.
I'm an AI Full-Stack Engineer working across the LLM stack — post-training, agents, retrieval, and the engineering that puts them into production. I build in the open and care about systems that hold up outside the demo.
Agentic RL
GRPO, multi-turn tool use, sandboxed execution, and execution-verified rewards.
See the reproductionMulti-Agent Systems
Orchestration, collaboration patterns, tool routing, and workflow engines.
View the frameworkNorway
Open to collaboration and AI engineering opportunities.
Connect on GitHub02 — Focus
Where I work best.
Agentic RL & Post-Training
Training agents that can reason, use tools, and verify their own work.
Multi-Agent Systems
Production-grade orchestration with collaboration modes, tool routing, and workflow engines.
Vertical-Domain LLMs
Domain adaptation from pretraining to SFT, RLHF, and quantization — starting with Chinese finance.
Applied NLP & Retrieval
BERT-family models, semantic matching, distillation, prompt learning, and Faiss retrieval.
03 — Selected Work
Open source, with receipts.
A few projects that show how I think about models, agents, and the systems around them.
Flagship · Open Source
Cornucopia-LLaMA-Fin-Chinese
An open-source Chinese financial LLM series with a lightweight vertical-domain training framework covering pretraining, SFT, RLHF, and quantization.
Production-grade multi-agent orchestration: 6 collaboration modes, 8 specialized tools, and 4 execution paths. Built with FastAPI and Streamlit.
End-to-end reproduction of MiMo-V2.6's interactive Agent RL pipeline on MS-Swift: multi-turn tool use, sandboxed execution, real-execution verification, and GRPO.
Enterprise-grade multi-agent travel assistant built from scratch on AgentScope 2.0.
TensorFlow 2.x architecture for loading BERT and its variants.
Prompt-learning algorithms for NLP tasks including text classification and information extraction.
TensorFlow 2.x toolkit for text similarity, knowledge distillation, interaction models, and dual-tower SBERT.
04 — Stack
Tools I reach for.
LLM & Training
Core model work
Python
PyTorch
TensorFlow
Hugging Face
MS-Swift
GRPO
Agents & Retrieval
Orchestration + context
AgentScope
LangGraph
MCP
RAG
Milvus
Neo4j
Engineering
Shipping it
FastAPI
Docker
Redis
Git
Linux
05 — GitHub
The work is public.
Open source, in the open
Open-source work with 736 stars and counting.
From a 650+ star financial LLM series to agent RL reproductions and the foundations of applied NLP — most of what I know lives in public repositories.
736
Total stars
25
Public repos
26
Followers
2017
Building since
Let's build
Working on agents, post-training,
or vertical-domain models?
I'm always glad to compare notes with people building real systems.