Visual resume + six-page portfolio

Yixin Zhang

张艺馨

R&D Engineer — Robotics infrastructure, applied ML, and LLM workflows.
This deck keeps the speculative notebook in the background: two pages of context, then six pages of concrete proof from real projects and volunteering.

Conversation starter

A systems-minded builder moving toward embodied intelligence

Current focus

Runtime infrastructure at the messy robot boundary

I work where sensors, OS primitives, robot middleware, and agent tools meet: camera/depth stream discovery, cross-view time sync, memfd IPC, ROS2 topics, C++ SDK design, CLI/REST interfaces, and AI-assisted app building.

Technical range
  • C++, Python, C#, Matlab; Linux and ROS2
  • Computer vision, XGBoost/LightGBM, deep reinforcement learning
  • RAG, agent skills, vector databases, multi-agent patterns
Robotics ML systems LLM agents
Education
4.61/5.0

NUS-ISS Graduate Diploma in Systems Analysis; HIT B.Eng. in Intelligent Test & Control Engineering.

Proof of work
24k+ stars

Core contributor to LLM-Cookbook; PRs merged into CAMEL and FastGPT for vector database work.

Filtered from notes

The original deck over-indexed on open questions. I kept only the useful conversation signal: perception, error recovery, data collection, agent memory, and runtime architecture.

Route map
2020–2024

Control engineering foundation at HIT; scholarship, innovation projects, deep RL capstone.

2023–2025

NUS-ISS systems training; open-source LLM education and security evaluation work.

2024–2025

A*STAR internship on climate mitigation analytics and GenAI-assisted case conversion.

2025–now

Robotics runtime engineering: media I/O, depth-camera quirks, ROS/DDS control, agent-assisted tooling.

Portfolio page 1 / 6

Five achievements selected from project experience and volunteering

01 · Computer vision

Video text extraction system

Deblurring + recognition pipeline for fast-moving, highly blurred text in video.

1st prize · selected national-level · Android deployment context
02 · Reinforcement learning

Delay-aware DQN path planning

Model-free dynamic navigation with real-time obstacle avoidance.

~12% faster convergence · +5.9% success rate
03 · LLM evaluation

SparkDesk KBQA red team

Agricultural RAG knowledge base plus adversarial prompts and jailbreak testing.

100+ QA cases · 10+ attacks · security analysis
04 · Open knowledge

LLM-Cookbook core contributor

Localized, standardized and tested one of the earliest influential Chinese prompt-engineering tutorials.

24k+ stars · reproducible prompts · interactive demo work
05 · Civic work

Volunteering and public science

Coastal cleanup, field research on industrial capability, and STEM translation.

5 tons team cleanup · key incubation team · Veritasium translation
Portfolio page 2 / 6
01 — COMPUTER VISION SYSTEM

Video Text Extraction System based on CNN and Attention

A team project from HIT's Innovation and Entrepreneurship Programme, built for extracting text from video when single-frame OCR fails because of motion blur, low resolution, and fast scene changes.

  • Architected a four-stage pipeline: text detection, frame tracking, multi-frame deblurring, and final text recognition.
  • Designed the deblurring module with adjacent-frame temporal information and attention mechanisms, inspired by BasicSR.
  • Proposed feature-level fusion to avoid image reconstruction loss and preserve stroke details before recognition.
  • Coordinated the project roadmap with Gantt planning and kept roughly 90% adherence to the planned schedule.
Image placeholder · replace with assets/portfolio/video-text.jpg
Pipeline diagram, blurry frame examples, recognition output
Recognition lift
+0.23% F1 vs SVTR-Tiny
Award
1st prize
Deployment
Old Android devices
Portfolio page 3 / 6
02 — DEEP REINFORCEMENT LEARNING

Delay-aware DQN for dynamic path planning

My HIT capstone explored how model-free path planners behave when observation updates are delayed, a setting closer to real dynamic environments than perfect synchronous simulation.

  • Enhanced a Deep Q-Network by introducing delayed observation updates for real-time obstacle avoidance.
  • Engineered goal-directed reward shaping with heuristics, then evaluated across sparse, dense, and zigzag obstacle grids in Gym.
  • Benchmarked against Rainbow, D3QN, DWA, and RT-RRT* to compare learning stability, computation cost, and path efficiency.
  • Built the project as a bridge between control intuition, simulation experiments, and ML evaluation discipline.
Image placeholder · replace with assets/portfolio/dqn-path-planning.jpg
Grid maps, learning curves, policy rollout traces
Convergence
~12% faster
Success rate
+5.9%
Baselines
Rainbow · D3QN · DWA · RT-RRT*
Portfolio page 4 / 6
03 — LLM EVALUATION & SECURITY

iFlytek SparkDesk KBQA: performance and attack surface

Instead of only asking whether a model could answer, this project tested how it answered under domain grounding, closed and open-ended questions, and adversarial pressure.

  • Crafted a domain-specific agricultural knowledge base to augment SparkDesk with retrieval.
  • Formulated 100+ closed- and open-ended questions and mixed in prompt injection plus jailbreak attacks.
  • Worked under Prof. Jie Liu's supervision to organize a team-based attack-and-defense evaluation process.
  • Produced a Chinese report connecting accuracy, vulnerability patterns, and KBQA design lessons.
Open project report
Image placeholder · replace with assets/portfolio/sparkdesk-kbqa.jpg
Question taxonomy, attack examples, result matrix
Test set
100+ QA cases
Security
10+ attacks
Mode
RAG + red team
Portfolio page 5 / 6
04 — OPEN-SOURCE EDUCATION

LLM-Cookbook core contributor

This was not just translation. The work made a fast-moving English tutorial usable for a Chinese developer audience by cleaning terminology, API behavior, formatting, and runnable examples.

  • Proofread the Chinese translation of DeepLearning.AI's ChatGPT Prompt Engineering for Developers.
  • Unified terminology, formatting, and API use across chapters after team discussion.
  • Validated localized prompts for reproducibility and engineered supplementary experiments.
  • Built extra interactive demo material for the “Chatbot” chapter.
Open LLM-Cookbook
Image placeholder · replace with assets/portfolio/llm-cookbook.jpg
Chapter screenshots, prompt tests, demo walkthrough
Audience
24k+ stars
Role
Core contributor
Artifact
Reproducible prompts
Portfolio page 6 / 6
05 — CIVIC PRACTICE & PUBLIC KNOWLEDGE

A wider proof of initiative: fieldwork, coastal cleanup, and science translation

These projects show how I work outside conventional engineering settings: joining physical labor, organizing research across cities, and helping make STEM content accessible to more people.

“Conversation starts faster when the proof is concrete.” This portfolio is designed around artifacts, not self-description.
Beach cleanup · Singapore

SG Beach Warriors

Worked on coastal conservation across North and East shorelines; as teams of 28 and 63, helped clean more than 5 tons of marine trash at North Coast.

Team laborEnvironmental action
Field research · HIT

“Exploring the Pillars of a Great Power”

Led a group of nine to research the Space Museum, World 5G Convention 2022, regional industrial enterprises, and museums in members' hometowns; the team won 2nd prize and was selected as a key incubation team.

Science communication

Veritasium translation

Contributed Chinese subtitle translation to Veritasium's official Bilibili channel, helping rigorous STEM concepts reach a broader audience.

View translated video
Image placeholder · replace with assets/portfolio/volunteering.jpg
Cleanup photos, field-research group, translation screenshot
1 / 8