Ziqian Bi 毕梓仟

Research Interests

Technical Skills

Industry Experience

China Post (Internship)

Jan 2026 - Mar 2026

AI Algorithm Engineer, Large Language Model Direction

GitHub link: chinapost-internship

• Built a China Post customer-service intelligent assistant with Django, django-ninja, Server-Sent Events (SSE), Bootstrap 5, and Tailwind CSS for the Web/API and streaming-response layer; integrated PostgreSQL + pgvector for vector storage and FAISS for local vector retrieval/indexing; used Ollama only as a local test backend for validating the LLM API adapter, with support for switching to vLLM, SGLang, OpenRouter, or OpenAI API backends for domain dialogue and ticket-generation workflows.
• Because much of the original project work was completed in closed internal systems, the public GitHub repository is a later reproduction built to show the architecture, API flow, and deployable demo without exposing internal data or infrastructure.
• Cleaned and expanded CSDS general customer-service data, filtered 5,376 postal multi-turn dialogues with a semantic-embedding and gpt-oss-20b multi-class Agent pipeline, and constructed 8,217 postal instruction-tuning samples after data augmentation and EDA analysis.
• Compared Qwen2.5-3B/7B, Qwen3, Llama 4, DeepSeek-V3/R1, gpt-oss-20b, and other open-source models across Chinese customer-service fit, JSON-structured output, small-data trainability, and deployment cost; selected Qwen2.5-7B-Instruct for SFT and used gpt-oss-20b as a JSON AI Agent tool for standardized ticket generation.
• Designed a reusable dual-task architecture for dialogue generation and ticket generation, separated functions through system-prompt switching, and fine-tuned with 4-bit quantized LoRA for a small GPU budget, reducing training time to under 10 seconds and preserving generalization with a three-level anti-overfitting mechanism.
• Implemented a cloud-native, containerized tiered-trigger RAG knowledge base for postal domain knowledge, using PostgreSQL + pgvector as the vector-storage layer and FAISS / SQLite for local retrieval and indexing workflows; the system was designed from the start for direct cloud deployment and supported high-risk query retrieval under 200 ms, JSON validation, batch export, and enterprise-backend integration.
• Accelerated system inference with torch.compile, pre-cached prefix KV cache, and context caching, reaching roughly 1.2 s dialogue inference and 1.7 s ticket generation while keeping output structure stable.
• Established a business-semantics evaluation protocol: single-turn dialogue accuracy reached 92.3%, multi-turn context understanding 87.6%, core ticket-field extraction 93.5%, hallucination rate 2.8%, expected labor cost reduction 30%, and ticket-triage efficiency improvement 80%.

Infraeo (Internship)

May 2025 - Jan 2026

Machine Learning and Deep Learning Engineer

GitHub link: Industrial-Query-Agent

• Many production details from the Infraeo internship were developed in closed internal systems, so the linked Industrial Query Agent repository is my later self-built reproduction that demonstrates the retrieval architecture, API design, and deployable service flow without exposing private code or data.
• Deployed and fine-tuned large language models up to 1000B parameters, including Qwen3 235B, Qwen3 Coder 480B, Kimi v2 1000B, Llama 3.3 70B, and GPT-OSS 120B, for domain-specific question answering with CoT prompting and RAG, reaching 89% answer accuracy.
• Built semantic retrieval over 2M+ documents with FAISS-based embeddings that generated 768-dimensional vectors and a BERT reranker, achieving recall@10 of 0.94 and precision@5 of 0.87.
• Fine-tuned Qwen and Llama models with LoRA and QLoRA using PyTorch, Unsloth, and Hugging Face Transformers, delivering 3.2x faster inference and 60% lower memory usage; developed backend-first AI services and API layers with FastAPI, Django, Docker, and Kubernetes supporting 500+ queries per second.
• Designed end-to-end machine learning pipelines processing 100GB+ of daily data, integrated vector databases (FAISS, ElasticSearch), PostgreSQL, and LangChain / LangGraph for multi-agent workflows, and packaged the reproduced system so it can be deployed directly to GCP or AWS.
• The frontend UI was not my main responsibility in the Infraeo work; my contribution focused on backend APIs, retrieval services, cloud-native packaging, and deployment architecture. In the reproduced public repository, the Django-template UI is only a lightweight demo wrapper around the backend service.

Chinese Academy of Sciences, Institute of Electrical Engineering (Internship)

Sep 2017 - Nov 2017

Smart Systems Solutions and Development Engineer

GitHub link: RS485-Modbus-Concrete-Sensor-Monitor

• Worked on an industrial-grade large-scale sensor-network monitoring system designed for major concrete infrastructure such as dams and bridges, using reliable RS485 vibrating-wire sensors and MODBUS RTU / TCP communication to collect field measurements through gateway-oriented industrial control flows.
• The original CAS implementation stayed on closed institute machines and could not be copied out, so the linked repository is a later self-built reproduction that preserves the architecture, data flow, and monitoring behavior without exposing internal code or deployment details.
• Designed the reproduced data path so simulated testing and production acquisition share the same backend contract: mock_server.py --feed writes 320 Redis keys per second for 20 gateways x 16 sensors with TTL=10s under monitor:sensor:{ip}:{index}, and Django RedisReader consumes the same keys for live frontend display; at the demonstrated update rate, the backend pipeline corresponds to tens of millions of sensor updates per day.
• For the production path, RS485 sensors feed collector/modbus_client.py, which writes the same Redis key format used by the mock flow; Redis handles most real-time caching and aggregation work as a high-performance layer, while the distributed Modbus gateway and acquisition-worker design is theoretically scalable to thousands to tens of thousands of sensors by adding more gateway devices and collector instances.

Research Experience

Education

AI-Augmented Engineering Scope

This is not Vibe Coding or casual exploration. It describes what I can build with Agentic AI tools through Test-Driven Development (TDD), Spec-Driven Development (SDD), Verification, reproducible implementation, Debugging, and maintainable production-grade engineering. The Technical Skills section is a conservative baseline for technologies I can understand, inspect, debug, modify, and maintain directly; with Agentic AI assistance, the practical implementation scope becomes broader, while my role shifts toward Specification, Verification, Integration, and Engineering Review rather than memorizing every implementation detail.