Professional Summary
Agentic AI Researcher and founder of LocoreMind, specializing in local agent models and LLM benchmarks. Developed LocoOperator-4B (17K+ downloads, HF trending top 6) and Eva-4B-V2 (84.9% SOTA, surpassing Claude Opus 4.5 and Gemini 3 Flash). Created EvasionBench, selected as official HuggingFace benchmark case (trending top 4). Published at ICLR 2026 and ICRA 2026. Expertise in knowledge distillation, multi-model consensus evaluation, and zero-cost local agent deployment. Track record of high-impact open-source contributions with 400+ GitHub stars and 20K+ HuggingFace downloads across projects.
Professional Experience
- Founded LocoreMind community for local agent model family, specializing in small models for private agentic workflows running on personal hardware
- Developed LocoOperator-4B (288 likes, 17K+ downloads, HF trending top 6, 139 GitHub stars) and LocoTrainer-4B (164 likes, 1.2K+ downloads, HF trending top 8, 99 GitHub stars) via knowledge distillation from Qwen3-Coder-Next
- Released qwen3.5-27b-cli-reasoning-3632x training dataset (52 likes, HF trending top 7) for agent research community
- Enabled zero-cost local agent deployment via llama.cpp for multi-turn codebase exploration and tool-calling
- Conducted research under the supervision of Prof. Yi Yang on financial communication analysis
- Developed Eva-4B (81.3% accuracy) and Eva-4B-V2 (84.9% SOTA, surpassing Gemini 3 Flash and Claude Opus 4.5) for detecting evasive answers in earnings call Q&A
- Created EvasionBench with 30K training samples using multi-model consensus and LLM-as-Judge protocol; selected as official HuggingFace benchmark case (117 likes, trending top 4, 33 GitHub stars)
- Published arXiv paper (2601.09142), featured on HF Daily Paper with 4K+ downloads and 95 likes
- Conducted research under the supervision of Prof. Yan LIN and Prof. Hongchuan SHEN on LLM applications and agent systems
- Created DramaBench benchmark for drama script continuation evaluation (110 likes, 490+ downloads on HuggingFace, published at ICLR 2026 Workshop)
- Developed multiple agent systems: Financial Agent, Computer Agent, ShortStudio video platform, and multimodal emotion analysis for drama content
- Fine-tuned Qwen3-8B for stance detection (0.77→0.93 accuracy) and open-sourced MSJ-Factory fine-tuning guide (150+ GitHub stars)
- Conducted research under the supervision of Prof. Dimitrios Kanoulas and Dr. Jianhao Jiao on embodied AI and robot navigation
- Implemented real-time object tracking and position-based planning on Unitree Go2 quadruped robot (published at ICRA 2026)
- Developed environmental mapping system integrating knowledge graphs, scene graphs, VLMs, and graph-based RAG for indoor navigation
- Developed embodied AI system for short video platform operations and RoboAgent hardware-software integration
- Built unmanned surface vessels for water photography, 3D reconstruction, and obstacle avoidance with depth estimation
- Conducted research under the supervision of Prof. Zeng Jicheng and Prof. Lin Yan on LLM applications in content analysis
- Collected 330K game reviews and built synthetic dataset for automated game issue detection using LLMs
- Applied difference-in-differences methodology to study LLM impact on HuggingFace ecosystem
- Developed TCMFormer Traditional Chinese Medicine diagnostic model and business data visualization systems
- Conducted machine learning research on public datasets, published multiple papers on prediction and classification tasks
- Collaborated with Prof. Xiao Li's robotics team at Huazhong University of Science and Technology on USV systems
- Developed autonomous navigation for rescue USV and implemented multi-agent reinforcement learning on vessel arrays
Technical Skills
Agent Frameworks:
LangChain, AutoGPT, BabyAGI, CrewAI, LangGraph, Semantic Kernel
LLM & APIs:
GPT, Claude, Gemini, Qwen, Mistral, OpenAI API, Anthropic API
Programming:
Python, TypeScript, JavaScript, C++
ML/DL Frameworks:
PyTorch, TensorFlow, Transformers, TRL
Vector Databases:
Pinecone, Weaviate, Chroma, Qdrant, FAISS
Tools & Platforms:
Docker, Kubernetes, AWS, GCP, FastAPI, Redis, PostgreSQL
Publications
- Shijian Ma, Yunqi Huang, Lin Yan. "I Can't Believe LLMs Still Can't Write Drama: Multi-Dimensional Failures in Script Continuation." ICLR 2026 Workshop ICBINB, 2026.
- Shijian Ma, Yan Lin, Yi Yang. "EvasionBench: Detecting Evasive Answers in Financial Q&A via Multi-Model Consensus and LLM-as-Judge." arXiv, 2026.
- Qianyi Zhang, Shijian Ma, Boyi Liu, Jianhao Jiao, Dimitrios Kanoulas. "Follow Everything: A Leader-Following and Obstacle Avoidance Framework with Goal-Aware Adaptation." ICRA, 2026.
- Shijian Ma, Shicong Ma, Jianhao Jiao. "RoboAgent: An Integrated Hardware-Software Agent for Short Video Interactions." TechRxiv, 2024.
- Shijian Ma, Shicong Ma, Jianhao Jiao. "WaveDepth: Enabling Water Depth Estimation with Water Segmentation." TechRxiv, 2024.
- Shijian Ma, Shicong Ma. "WaveShot: A Compact Portable Unmanned Surface Vessel for Dynamic Water Surface Videography and Media Production." arXiv, 2024. AIME accept.
- Shijian Ma, Shicong Ma. "Exploring the Relationship between Vehicle Specifications and Fuel Efficiency using Random Forest Regression and Gradient Boosting Regression Models." ECCS, 2024.
- Shijian Ma. "Churn Prediction in Business using Logistic Regression and Logit Boost." ICDSCA, 2023.
- Shicong Ma, Xiaoguang Tang, Shijian Ma. "Integration of Multi-sensor Marine Environment Monitoring System for Hybrid-Power USV." ICGMRS, 2023.
- Shijian Ma, Xiaoguang Tang, Weize Ma, Jitao Liu, Li Xiao. "An Unmanned Cleaning Robot for the Inner Wall of Large LNG Tanks." ICARM, 2022.
Key Projects
- Founded open-source community for local agent models designed for private agentic workflows running on personal hardware
- Developed LocoOperator-4B: 4B-parameter tool-calling agent trained via knowledge distillation from Qwen3-Coder-Next for multi-turn codebase exploration
- Created LocoTrainer-4B and LocoTrainer framework for agent training workflows with zero API cost via llama.cpp
- Released qwen3.5-27b-cli-reasoning-3632x dataset for agent training research
- Achieved significant community impact: LocoOperator (288 likes, 17K+ downloads, HF trending top 6, 139 GitHub stars), LocoTrainer (164 likes, 1.2K+ downloads, HF trending top 8, 99 GitHub stars), Dataset (52 likes, HF trending top 7)
- Developed Eva-4B (81.3% accuracy) and Eva-4B-V2 (84.9% SOTA, surpassing Gemini 3 Flash and Claude Opus 4.5) for detecting evasive answers in earnings call Q&A
- Created EvasionBench with 30K training samples via multi-model consensus; selected as official HuggingFace benchmark case
- Community impact: Eva-4B-V2 (113 likes, 1.1K+ downloads, trending 21), EvasionBench (117 likes, trending top 4, 33 GitHub stars), Paper (4K+ downloads, 95 likes, HF Daily Paper featured)
- Fine-tuned Qwen3-4B-Instruct-2507 for detecting evasion levels in earnings call Q&A responses
- Implemented Rasiah taxonomy-based classification system with three evasion categories
- Achieved community adoption with 36 likes and 145 downloads on HuggingFace
- Developed first large-scale benchmark for drama script continuation with six-dimensional evaluation framework
- Evaluated 8 SOTA models on 1,103 scripts (8,824 evaluations); published at ICLR 2026 Workshop ICBINB
- Community impact: 110 likes, 490+ downloads on HuggingFace, 78 GitHub stars, 16 Daily Paper upvotes
- Full parameter fine-tuned Qwen3-8B on 6,319 dramatic script samples using ms-swift framework
- Achieved +262% output length and +80% thinking depth with explicit creative reasoning chains
- Deployed on HuggingFace achieving trending rank #28 with 89 likes and over 2,000 downloads
- Financial Agent: Automated financial text analysis with continuous learning for evolving data
- XAgent: X.com automation framework using Claude Agent SDK and Chrome DevTools MCP (26 GitHub stars)
- Computer Agent: Natural language computer automation with screen capture analysis and multi-step task planning
- ShortStudio: AI-powered video creation platform analyzing trending content and generating similar videos
- Multimodal Emotion Analysis: Drama content analysis combining visual, audio, and textual features
- Created comprehensive guide for fine-tuning Qwen2.5-Coder for Chinese sentiment analysis
- Fine-tuned Qwen3-8B model for stance detection using Zhihu data
- Achieved significant accuracy improvement from 0.77 to 0.93 in stance detection tasks
- Open-sourced project gaining 150+ GitHub stars for community contribution
- Implemented real-time object tracking and position-based planning on Go2 quadruped robot
- Achieved open-set object following capabilities for diverse target objects
- Created maps by integrating predefined knowledge graphs with real-time observations
- Integrated scene graphs, vision-language models, and graph-based RAG
- Applied to indoor navigation and scene understanding tasks
- RoboAgent: Autonomous social media interaction system with computer vision and LLM-generated content (published at TechRxiv)
- WaveShot: Portable USV for water videography with monocular depth estimation (published at arXiv, AIME accepted)
- WaveDepth: Water depth estimation combining semantic segmentation with specialized dataset (published at TechRxiv)
- Underwater Perception: 3D reconstruction system integrating sonar, RTK-GPS, and point cloud visualization
- Patrol Boat: Real-time streaming USV with 4G module, LLM integration, and generative AI capabilities
Honors & Achievements
Professional Service
- Reviewer: ICLR 2026, IROS 2025, ICRA 2025, RA-L 2024, IROS 2024
Patents
- ZL201420713314.1 - Portable remote control water cleaning boat (2022)
- ZL202130395091.4 - Water surface intelligent cleaning robot (2022)
Awards
- Silver Medal, 7th National College Students "Internet plus" Entrepreneurship Contest (2022, Hubei Province)
- Winner's Prize, 16th "Chunhui Cup" Chinese Overseas Students Innovation Competition (2022)