Guangchi Yuan — Systems Architect for Autonomous Intelligence

Identity

Profile

Role

Systems Architect
Tech Lead
Engineering Leader

Domain

Autonomous Driving
Industrial Intelligence
AI Architecture

Focus

System Decomposition
Multi-layer Memory
Abstract to Deploy

I am not a single-point technologist. From autonomous driving perception systems to industrial AI infrastructure, from 3D LiDAR SLAM to multi-layer Agent memory architectures, my trajectory spans multiple domains — yet the underlying thread remains constant: how to enable machines to build reliable understanding, memory, and decision-making loops under real-world constraints.

At Haibo Intelligent Technology, I built a perception and fusion-localization system for 80-meter mining trucks operating under full working conditions from scratch, leading a 16-person team through 100,000+ full-cycle validations. Earlier, through an independent venture, I practiced 3D LiDAR SLAM, graph optimization, and multi-sensor fusion localization hands-on — learning firsthand the gap between algorithm prototypes and production systems. My NLP background provided a solid foundation in sequence modeling, semantic representation, and large-scale data processing — all of which became essential for understanding LLMs and Agent systems.

Today, my focus has expanded to the frontier: how long-term memory in AI systems is organized and evolves, how Agent state spaces are designed, how belief and policy are decoupled and updated, and how orchestration and execution governance work in complex task systems. The next generation of AI will not be larger models — it will be systems with structured memory, verifiable execution, and continuous evolution.

My habit is to dig deep into the representation layer: is the coordinate system unified? Are numerical channels designed correctly? Is structural prior properly encoded? How does error propagate through the system? This habit grounds architectural decisions in clear understanding of system structure, not intuition.

Core Domains

Domains

Seven technical directions organized by domain. Not a skill list — different prisms for understanding complex intelligent systems.

01

Autonomous Systems & Robotics

Autonomous Systems

Built an unmanned mining truck perception system from scratch, covering LiDAR, mmWave radar, and vision multi-modal fusion, supporting 80-meter trucks in dust, vibration, and lighting extremes. Led a 16-person team through 100,000+ full-cycle validations with fault rate below 0.01%.

02

Perception Fusion & Spatial AI

Perception & Spatial AI

Multi-sensor fusion, object detection and tracking, OCC occupancy networks, HD map crowd-sourced updates. Spatial intelligence is the foundation of physical-world AI — precise perception and localization are prerequisites for all higher-level decisions. From point clouds to images, from local features to global maps.

03

Localization & State Estimation

Localization & Estimation

Deep hands-on experience with EKF / ESKF / graph optimization / GTSAM / LOAM series, achieving large-scale real-time mapping and high-precision indoor/outdoor localization. Fusing IMU, RTK, wheel odometry, and designing 100+ fault-response strategies to ensure localization continuity in extreme scenarios.

04

AI Architecture & Agent Systems

Agent OS & Memory

Exploring multi-layer memory system design — working memory, episodic memory, semantic memory with promotion/demotion mechanisms and long-term context governance. Focused on decoupling Memory / State / Policy, belief graphs with Bayesian updates, enabling Agents to reason from structured belief rather than context windows alone.

05

Geometry & Numerical Representation

Geometry & Representation

Is the coordinate system unified? Does normalization destroy physical meaning? Is there aliasing in feature space? These questions extend to CAD/CAM intelligence, STEP-to-G-code structured intermediate representation. Correct intermediate representations determine system extensibility; numerical modeling must match geometric structure.

06

Engineering Reliability & Safety

Reliability & Safety

Safety is not an add-on — it is part of the system's ontology. Redundant architectures, fault detection, and graceful degradation must be considered from day one. From sensor redundancy to algorithm redundancy, from fault detection to degradation paths, every link's reliability determines survival in real-world scenarios.

07

Research Insight to Deployed Systems

Research to Deployment

Full-cycle 0-to-1 capability: algorithm design, engineering, performance optimization, risk control, team leadership. From demo to long-running systems — it takes more than technical skill. It requires clear vision for system evolution paths and deep understanding of real-world constraints.

Trajectory

Path

From NLP to autonomous driving perception and mapping, to system architecture and AI work-system design. Not scattered skills — a deepening understanding of complex intelligent systems.

2021.08 — Present

Perception Development Manager

Anhui Haibo Intelligent Technology Co., Ltd.

Built the unmanned mining truck perception system from scratch, establishing a multi-layer redundant perception architecture covering LiDAR, mmWave radar, and vision multi-modal fusion. Led a 16-person team through the full pipeline from algorithm R&D to production deployment, supporting 80-meter trucks in extreme working conditions around the clock. Owned AEB, object tracking, trajectory prediction, sensor calibration, HD mapping, crowd-sourced updates, and perception fault diagnosis.

99.9% detection @ 80m 100+ fault strategies 100K+ cycle validations <0.01% fault rate 16-person team Multi-modal redundancy

2019.11 — 2021.08

Co-Founder

Yuejie (Shanghai) Technology Co., Ltd.

Independently explored 3D LiDAR SLAM and high-precision localization, gaining deep hands-on experience with LOAM-series algorithms, graph optimization, IMU fusion, and RTK integration. Achieved large-scale real-time mapping and indoor/outdoor high-precision localization. Explored PointNet point-cloud global localization, extreme-scenario recovery, and conducted applied research in ESKF, EKF, SSD, and related directions.

LOAM + Graph Opt IMU + RTK Fusion PointNet Global Loc Large-scale Mapping Extreme Recovery

2018.02 — 2019.11

NLP Algorithm Lead

Guangzhou Bailing Data Co., Ltd.

Built and managed the NLP algorithm team, leading bidding-data processing and public-opinion analysis systems. Drove early adoption of BERT and large models in production, building a pipeline processing 30M+ texts daily with 95%+ classification accuracy. This experience laid the foundation in sequence modeling and semantic representation that later became essential for understanding LLMs and Agent systems.

BERT Early Adopter 30M+ daily 95%+ accuracy Team Building

Frontier Research

Frontier

Ongoing thinking about the architectural form of next-generation AI systems. Not paper abstracts — active exploration directions.

AI-Driven Geometry & Manufacturing Intelligence

Exploring intelligent STEP-to-G-code conversion, structured intermediate representations in CAD/CAM. Making AI understand manufacturing semantics — not just generating geometry, but producing verifiable, simulable, executable structured data.

Agent OS & Multi-Layer Memory Systems

Designing OS-level Agent architecture: layered organization of working, episodic, and semantic memory; promotion/demotion mechanisms; long-term context distillation and governance strategies.

Memory / State / Policy Decoupled Architecture

Decoupling memory storage, state representation, and decision policy. Exploring belief graphs with Bayesian update mechanisms, enabling Agent decisions grounded in structured world belief rather than context windows alone.

Orchestration & Execution Governance

Worker-reviewer architecture, task decomposition and recomposition, execution verification and rollback mechanisms. Complex tasks are not single LLM calls — they are multi-stage, inspectable, repairable governance pipelines.

Industrial AI Closed-Loop Systems

From model output to structured execution systems: verifiable, simulable, feedback-driven, evolvable. Industrial AI must form a closed loop — perceive, understand, decide, execute, verify, learn, repeat.

// System Notes

True Agents are not prompt engineering — they are state-space and memory architecture engineering
The key to long-term memory is not storage, but retrieval relevance and synthesis capability
Belief updates must be traceable, explainable, and reversible
The core of orchestration is graceful degradation and state recovery on failure
The leap from chat tool to work system requires an execution layer, not just generation
The value of industrial AI is not in generation, but in structured accumulation and closed-loop verification
Intermediate representation quality determines the system ceiling
Models iterate; architectures endure
Bottlenecks in complex systems are usually at the interface layer, not the algorithm layer
Sustainable evolution beats one-time optimality

Capability Map

Structured capability matrix. Every group is backed by deep practice on real projects.

Autonomous Driving

Autonomous Systems

Multi-modal Fusion AEB Object Tracking Prediction HD Mapping Calibration Diagnosis Redundancy

Perception & Spatial

YOLO OCC PointNet PointNetVLAD SLAM ICP NDT LOAM

Localization & Estimation

EKF ESKF GTSAM Graph Opt IMU Fusion RTK Global Loc Registration

ML & Deep Learning

ML / DL

PyTorch TensorFlow CUDA Transformer BERT OpenCV Optimization Numerical

NLP / LLM / Agent

Language & Agent

NLP Pipeline LLM Apps Agent Memory Orchestration Prompt Eng Semantics Classification Sentiment

Engineering

C++ Python Linux x86/ARM Performance Deployment CI/CD Version Ctrl

Geometry & Repr.

Coord Systems Structure Prior Intermediate Repr. Error Propagation Closed-loop Numerical CAD/CAM STEP/G-code

Leadership

Team Building Tech Planning Project Mgmt Cross-functional 0-to-1 Mentoring Risk Mgmt Architecture

Operating Principles

Core principles guiding how I design complex systems. Not motivational quotes — engineering judgment criteria.

01

Complex systems must be layered, or they become unmaintainable

Each layer focuses on its own level of abstraction, interacting with adjacent layers through clear interfaces. Chaotic systems usually result from blurred responsibilities between layers and undisciplined data penetration. Layering is not over-engineering — it is the only effective way to control complexity.

02

Memory, state, and decision-making should not be mixed together

These three operate on different time scales and update frequencies. Memory accumulates over the long term; state changes in real time; decisions follow independent policy logic informed by the other two. Mix them, and the system collapses on at least one dimension.

03

A truly usable AI system must have a feedback loop

Perceive, understand, decide, execute, verify, learn — this loop must be complete. Break any link and the system degrades. In industrial settings especially: no feedback, no reliability.

04

Intermediate representations and interfaces determine system extensibility

The system's ceiling is not how advanced its algorithms are, but whether its intermediate representations are general enough and its interfaces stable enough. Good representations let modules evolve independently; bad ones make every upgrade a minefield.

05

Safety, redundancy, and fault strategy are part of the system's ontology

Not afterthought patches, not bolt-on modules. From day one of architecture design: when this module fails, how does the system continue? What is the fault-detection latency budget? Is the degradation path clear?

06

Production deployment is not a compromise on research — it validates whether the abstraction is correct

If an abstraction hits walls everywhere during deployment, the problem is usually not engineering — it is the abstraction itself. Real-world constraints are the best litmus test. A system that can ship has a correct architecture.

07

Great systems are not built by stacking features, but by clear boundaries, stable structure, and sustainable evolution

Features can be added incrementally, but structural rot is often irreversible. Guarding architectural boundaries matters more than shipping a feature. Great systems are organisms — they grow, heal, adapt. Not Jenga towers that get more fragile with every block.

08

Models iterate; architectures endure

Today's SOTA model may be surpassed tomorrow, but a clean system architecture can carry multiple generations of algorithms. Investing in architectural correctness pays more long-term dividends than chasing the latest model. Architecture is the skeleton; models are the flesh — the skeleton stands, the flesh can be swapped.

Education

NY

New York University

M.S. in Electrical Engineering

2015 — 2017

HH

Hohai University

B.Eng. in Electrical Engineering and Automation

2009 — 2013