Fenyong Li

Full-Stack Engineer (AI Engineer)

• Enterprise-Scale Agentic Workflow: Addressed fragmented content planning, writing, review, approval, and publishing processes with long manual turnaround times by leading the design and implementation of an end-to-end Agentic Workflow for an organization of 160,000+ users. Orchestrated multi-stage agent workflows with Human-in-the-Loop compliance controls, reducing content production cycles from weeks to hours.
• Production-Grade Agent Runtime and RAG: Addressed unstable reasoning, execution interruptions, unauthorized tool usage, and enterprise knowledge access constraints in long-running agent workflows by building a recoverable execution framework with LangGraph + PostgreSQL Checkpointing. Integrated dynamic Pydanticschemas,permission-awareTool-basedRAG,andprogressivecontextdisclosuretoreduceirrelevant context, improve multi-step reasoning stability, enforce controlled tool usage, and enable permission-based access to enterprise knowledge.
• Production Engineering and Cloud Deployment: Reduced regression and deployment risks in rapid Agent system iterationbydeveloping1,000+pytestunit,integration,andregressiontestsandbuildingaDocker-basedCI/CDpipeline using GitHub Actions → AWS ECR → ECS Fargate. Integrated RDS PostgreSQL + Alembic for production data persistence, schema evolution, and database migrations.

Technical Consultant (AI & Automation Solutions)

• Multi-AgentOrchestrationandReasoning: Addressedlimitedreasoningstabilityandtaskdecompositionincomplex decision-support scenarios by independently building an end-to-end AI decision-support system with LangGraph and adopting the ReAct framework for autonomous multi-step problem solving. Developed a modular tool library for the supervisor agent to dynamically route tasks to specialized sub-agents, reducing hallucinations and achieving 95%+ accuracy across iterative prompt evaluations.
• HybridRAGandRetrievalEnhancement: Addressedweakkeywordrecallandsemanticmatchingerrorsinenterprise knowledgeretrievalbybuildingahybridBM25+NeonpgvectorretrievalpipelinewithDashScopegte-rerankerfor second-stage reranking. Integrated SHA-256 deduplication and Ragas (LLM-as-a-Judge) evaluation to continuously optimize retrieval quality while processing 10,000+ data nodes per week.
• Full-Stack Streaming and LLMOps: Improved real-time interaction and production observability for long-running agentreasoningworkflowsbyimplementingstreaminginferencewithFastAPI + Next.js + SSEandcontainerizingthe applicationwithDocker. DeployedservicesacrossVercel,Render,andCloudflareR2,whileintegratingLangSmith for end-to-end tracing, debugging, and Token cost analysis and optimization.

System Engineer

• NALA+ System Development: Contributedsignificantlytotheend-to-enddevelopmentoftheNALA+IoTgateway, covering requirement analysis, architectural design, core control logic, and production delivery. Managed cross-platform hardware driver adaptation and low-level protocol stack implementation, ensuring system stability in complex overseas industrial environments.
• Communication Architecture & Optimization: Redesigned and implemented an asynchronous communication mechanism for distributed nodes to resolve network congestion and packet loss during high-concurrency sensor data ingestion. Conductedextensivestresstestingandperformancetuning,reducingsystemresponselatencybynearly30%.
• JTC Project Implementation: Spearheaded energy simulation and energy-saving strategy iterations for Singapore JurongTownCorporation(JTC)officebuildings. OptimizedlightingcontrolalgorithmstoreduceLightingPowerDensity (LPD)to5W/m²,helpingtheprojectachievethehighestenergyefficiencystandardssetbytheBuildingandConstruction Authority (BCA).

Architected an autonomous multimodal AI tutoring agent, leveraging Augmented Reality to establish real-time visual perception and spatial awareness of the user’s physical environment. Engineered a cognitive decision-making core using GPT,incorporatingChain-of-Thought(CoT)promptingtodynamicallyanalyzeuserinteractionsandautonomouslyformulate personalized pedagogical strategies. Designed a multi-tier memory architecture (short-term dialogue buffering and long-term user profile retrieval) to maintain context continuity and adapt to the user’s evolving learning curve. Optimized the Python-Unityintegrationpipelinetoinstantlytranslatetheagent’slogicaldecisionsintosynchronized,multimodalexecutions, including fluid voice output and corresponding 3D avatar animations.

Spearheaded a comparative evaluation between a discriminative CNN-LSTM baseline and an advanced generative BLIP frameworktodeveloprobustperceptionlayersforclinicalAIassistants. LeveragedKaggle’shigh-performanceP100GPUsand theHuggingFaceTransformersecosystemtoorchestratecomplexmultimodaltrainingpipelines. Engineeredavisual-linguistic reasoning core using Vision Transformers (ViT) and Cross-Attention mechanisms. Successfully overcame the traditional ”classification trap” via auto-regressive token generation, significantly improving the agent’s semantic reasoning (F1 Score) and zero-shot Out-of-Vocabulary (OOV) capabilities.

Architected a full-stack ”Edge-Cloud” system for precision agriculture. Integrated CV-based facial authentication and millisecond-level fire warnings. Developed a sensor-driven intelligent ”Fertigation” system. Synergized UGV autonomous inspection with fixed stations for full-dimensional monitoring, achieving closed-loop management. Awarded ”Outstanding Graduation Project”.

Designed a high-extensibility RPG core framework using Unity 3D. Developed a complex NPC Behavior Tree system and a dynamic quest-driven system to support flexible AI interactions and branching logic. Independently implemented a reliable Client-ServerStateSynchronization(StateSync)mechanismtoensuremulti-terminalconsistency. Optimizedsystemrobustness throughdecoupledinventoryandevent-drivenarchitectures. Recruitedasthefirstcorememberoftheuniversity’stop-tiertech society in recent years due to exceptional technical execution.

Bachelor of Science · Computer Science and Technology

Master · Computer Science