Building open-source LLM & agent systems — from post-training to production.
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.
- Building Cornucopia — an open-source Chinese financial LLM series and a lightweight vertical-domain training stack (pretraining → SFT → RLHF → quantization).
- Exploring agentic RL with GRPO, multi-turn tool use, sandboxed execution, and execution-verified rewards.
- Designing production-grade multi-agent architectures, collaboration patterns, and workflow engines.
- Based in Norway · open to collaboration and AI engineering opportunities.
| Area | Work |
|---|---|
| Agentic RL & Post-Training | GRPO, multi-turn tool use, sandboxed execution, and execution-verified rewards. |
| Multi-Agent Systems | Orchestration, collaboration patterns, tool routing, and workflow engines. |
| Vertical-Domain LLMs | Financial LLMs in Chinese, from pretraining and SFT through RLHF and quantization. |
| Applied NLP & Retrieval | BERT-family models, semantic matching, distillation, prompt learning, and Faiss retrieval. |
- Cornucopia-LLaMA-Fin-Chinese
— An open-source Chinese financial LLM series with a lightweight vertical-domain training framework covering pretraining, SFT, RLHF, and quantization.
- Cornucopia-Multi-Agent
— Production-grade multi-agent orchestration: 6 collaboration modes, 8 specialized tools, and 4 execution paths, built with FastAPI and Streamlit.
- mimo-agent-grpo-swift
— 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.
- aligo-travel-agents
— Enterprise-grade multi-agent travel assistant built from scratch on AgentScope 2.0.
- xbert
— TensorFlow 2.x architecture for loading BERT and its variants.
- Semantic-Similarity-Match
— TensorFlow 2.x toolkit for text similarity, knowledge distillation, interaction models, and dual-tower SBERT.
- NLP-Prompt
— Prompt-learning algorithms for NLP tasks including text classification and information extraction.
More open-source work
- Agentic-AI-Knowledge — Working notes on LangGraph, LightRAG, Milvus, Neo4j, Claude Agent SDK, and multi-agent architecture.
- Faiss-FAQ_roboat — Applying Faiss to FAQ retrieval and question answering.
- Text-CLS — TensorFlow 2.x text classification with adversarial training, label smoothing, and class-imbalance handling.
- Entity-Extraction — Entity extraction with PyTorch / TensorFlow 2.x.
- Property-Graph-by-python3.5 — A property graph implementation in Python.
LLM & Training
Agents & Retrieval
Engineering
If you're working on LLM agents, post-training, or vertical-domain models, I'd be glad to compare notes.
Thanks for stopping by.