AI Research · Systems · Product
Chening Yang
Building Neural-Symbolic Agentic Systems and advancing Model Domain Adaptation
Building on prior work in VRDU and Agentic RAG, current focus over the last year:
- Neural-Symbolic Agentic Systems
- Model Domain Adaptation
- Production LLM Serving
02 · Expertise
Core Competencies
AI Research & Development
Leading innovation in:
- Specialist Model Domain Adaptation
- Master Large Language Models (LLMs)
- Proficient Document Understanding & Multimodal AI
Technical Leadership
Driving excellence through:
- Lead Cross-functional R&D Team Management
- Principal Project Architecture
- Strategic Roadmap Planning
System Architecture
Specialized in:
- Master Neural-Symbolic Agentic Systems
- Advanced Agentic RAG Systems
- Proficient Cloud Infrastructure
Product Development
End-to-end expertise in:
- Lead AI Product Strategy
- Advanced Solution Design
- Skilled Production Deployment
03 · Connect
Let's Connect
$ cat contact_details.txt
04 · Research Log
Research Timeline
2026 H1
Model Finetuning & Domain Adaptation
Focused on domain adaptation pipelines spanning CPT, SFT, and GRPO, including joint training strategies, catastrophic-forgetting mitigation, and model weight merging.
Tech Stack: Unsloth, TRL, TorchTune.
Model Serving & Inference Optimization
Advanced production serving with quantization and speculative decoding to improve throughput and latency for LLM inference.
Tech Stack: SGLang, vLLM, Eagle, NEXTN, bitsandbytes, AWQ, GPTQ, FP8.
Data Cleaning & Generation
Built data cleaning and synthetic data generation workflows to support high-quality finetuning corpora.
Agent Harness & LLM Gateway
Developed agent harnesses and unified LLM gateway layers for reliable multi-provider orchestration and agentic execution.
Tech Stack: DeepAgents, MetaGPT, LiteLLM.
2025 H2
Prover Models & Formal Language Systems
Explored prover models grounded in formal language systems to strengthen verifiable reasoning and theorem-proving capabilities.
Tech Stack: Lean.
Neural-Symbolic Agents
Built neural-symbolic agents that combine symbolic solvers with LLM planning for structured reasoning and tool use.
Tech Stack: Z3, Smolagents, LangGraph.
Reasoning Architecture & Instruction Finetuning
Researched reasoning architectures and instruction finetuning with a focus on improved tool calling for agentic workflows.
Tech Stack: Unsloth, Transformers, PEFT.
Small Model Serving & Chat Interfaces
Deployed small-model serving stacks and chat interfaces to support practical agentic architectures end to end.
Tech Stack: llama.cpp, Ollama, Open WebUI.
2025 H1
Production LLM / Embedding Serving
Shipped LLM and embedding serving stacks into production, focusing on throughput, latency SLOs, rollout reliability, and operational readiness for high-traffic applications.
Tech Stack: Text Generation Inference, vLLM, SGLang, Text Embeddings Inference.
Shipping Multi-Agent RAG
Productized agentic RAG systems for real user workloads, emphasizing stable tool calling, memory/state management, failure handling, and end-to-end workflow execution.
Tech Stack: MetaGPT, AutoGen, LangGraph.
LLM Ops & Observability
Built production monitoring and evaluation loops for LLM applications, covering tracing, quality regression checks, and model selection trade-offs for deployment.
Tech Stack: LangSmith, DeepSeek, LLaMA, Qwen.
Production Graph RAG Pipelines
Deployed graph-based retrieval pipelines into production, integrating indexing, query orchestration, and domain-entity graphs for reliable retrieval performance at scale.
Tech Stack: Microsoft GraphRAG, Neo4j.
2024 H2
LLM / Embedding Model Serving
Researched scalable and efficient methods for serving LLMs and embedding models in production environments to support high-throughput applications.
Tech Stack: Text Generation Inference, vLLM, SGLang, Text Embeddings Inference.
Multi-Agent RAG
Explored advanced agentic RAG frameworks for scalable tool usage, memory organization, query management, reflection, reasoning, and execution.
Tech Stack: MetaGPT, AutoGen.
LLM Architecture
Conducted a comprehensive survey of fundamental structures and design principles behind various LLM models.
Tech Stack: DeepSeek, LLaMA, Qwen, Titans.
Graph-Based RAG
Researched graph-based retrieval paradigms and algorithms to enhance retrieval performance, exploring single- and multi-graph architectures, as well as layout-based and domain-entity graphs.
Tech Stack: Microsoft GraphRAG, Neo4j.
2024 H1
Document Structurization
Built an end-to-end solution to transform unstructured or semi-structured data into structured formats easily interpretable by LLMs. Examples include converting Excel/PDF files into JSON or Markdown.
Tech Stack: Transformers, Gradio, Docling, Unstructured-IO, LlamaParse, Azure AI Document Intelligence.
LLM Reasoning and Acting
Researched various LLM reasoning schemas with Human-in-the-loop approaches, including Chain-of-Thought (CoT), Tree-of-Thought (ToT), Self-Consistency, ReAct, and Retrieval-Augmented Workflow Optimization (ReWOO).
Tech Stack: LangGraph, LangSmith, LlamaIndex, Haystack.
Retrieval-Augmented Generation
Developed an LLM-powered chatbot utilizing Retrieval-Augmented Generation (RAG) to effectively leverage user-provided data for generating accurate and contextually grounded answers.
Tech Stack: Elasticsearch, Milvus, Kotaemon.
Trustworthy LLM
Investigated methods to reduce LLM hallucinations and enhance knowledge grounding for trustworthy AI outputs.
Tech Stack: TruLens, Ragas, DeepEval.
Earlier research archive 2019—2023
2023 H2
Tech Stack Management
Investigated effective approaches to organize and manage technical assets for streamlined workflows.
Tech Stack: Hugging Face Hub, Hugging Face Spaces.
Table Detection and Recognition
Researched methodologies for extracting table information, exploring specialized models for table detection and recognition.
Tech Stack: DETR, TATR.
ETL Pipeline
Designed an efficient ETL pipeline for performance tracking and continuous improvement of workflows.
Tech Stack: Apache Airflow, Docker Swarm, Kubernetes.
LLM Training
Researched techniques for training and serving large language models (LLMs), covering pretraining, supervised training, preference alignment (RLHF), quantization, and low-rank adaptation methods.
Tech Stack: PEFT, OpenAI, bitsandbytes, W&B, TorchTune, LangChain, Accelerate.
Multi-Label Classification
Researched lightweight CNN and ViT architectures to develop efficient multi-label classifiers for diverse datasets and tasks.
Tech Stack: Dino, PyTorch Image Models.
2023 H1
Lightweight Multimodal Architecture
Developed a CPU-compatible transformer-based architecture for information extraction, integrating image and text inputs while emphasizing spatial information learning. Achieved state-of-the-art performance among lightweight transformer models. Published two accepted papers at ICDAR 2024.
Tech Stack: Transformers, FAISS.
Text Recognition
Researched advanced OCR architectures, focusing on transformer-based encoder-decoder models with innovative image synthesis and augmentation techniques.
Tech Stack: PyTorch Lightning, Albumentations.
Training Strategies
Experimented with training techniques for CNN and transformer-based architectures, including Stochastic Depth, Teacher-Student Knowledge Distillation, Label Smoothing, and learning rate warm-up strategies.
Tech Stack: PyTorch Image Models (timm).
2022 H2
Document Object Detection
Experimented with anchor-based and anchor-free architectures for document object extraction.
Tech Stack: Ultralytics, MMDetection.
Sentence Embedding
Researched lightweight multilingual sentence embedding techniques for improved efficiency and scalability.
Tech Stack: Sentence Transformers (SBERT).
2022 H1
Image Registration
Researched local image feature matching using OpenCV-based methods (e.g., SIFT, ORB) and deep learning solutions (e.g., LoFTR).
Tech Stack: OpenCV, Kornia.
Cloud Deployment
Deployed an Intelligent Document Processing (IDP) pipeline on AWS, experimenting with distributed system designs for scalability and performance.
Tech Stack: Docker, AWS Services, Seldon, Celery, FastAPI, Uvicorn.
2021 H2
Key-Value Pair Extraction
Explored transformer-based architectures for key-value extraction, focusing on BERT-based models and their variations.
Tech Stack: PyTorch, Hugging Face Transformers.
2021 H1
Model Compression
Researched Knowledge Distillation and Post-Training Quantization (PTQ) for lightweight, high-performance models.
Tech Stack: OpenVINO, TorchScript.
Model Serving
Investigated efficient model serving methods using ONNX and TensorRT for optimized deployment.
Tech Stack: TensorRT, ONNX.
2020 H2
Key-Value Pair Extraction
Applied DGCNN and Residual Gated Graph ConvNets for key-value extraction. Enhanced spatial learning with CenterNET and explored metric learning techniques for embedding and clustering.
Tech Stack: DGL, PyTorch Hub, pytorch-metric-learning.
Sentence Segmentation
Developed strategies using Named Entity Recognition (NER) to segment sentences into meaningful pieces. Implemented GRU/LSTM with CRF loss for tokenization and tagging.
Tech Stack: pytorch-crf.
2020 H1
Node Classification
Investigated graph construction and applied Graph Attention Networks (GANs) for document node classification tasks.
Tech Stack: DGL.
Document Embedding
Researched graph-based solutions, including Graph Convolutional Networks (GCNs), for effective document embedding and representation.
Tech Stack: DGL.
2019 H2
Text Detection
Researched CNN backbones and loss functions for encoder-decoder architectures in text detection.
Tech Stack: Sagemaker, S3, TensorFlow, imgaug.
Text Recognition (OCR)
Improved CRNN+CTC architecture for unlimited-length OCR decoding.
Tech Stack: Sagemaker, S3, PyTorch.
2019 H1
Person Re-Identification
Researched face detection and recognition solutions integrated with human re-identification architectures to develop a surveillance application.
Tech Stack: FaceNet, R-CNN, DeepFace, MTCNN, SSD.