# Kefei Yan

> Georgia Tech MSCS, building AI for the physical world. Graduate researcher, HAAG @ Georgia Tech · graduating December 2026. This file is the LLM-readable version of kefeiyan.com. Humans get an assistant interface at the homepage and the full plain profile at /classic.html; you get structured facts here.

## About

Currently pursuing an MSCS at Georgia Tech (AI track), graduating December 2026, and doing biodiversity-monitoring research with the HAAG group there. Previously completed an online MS in Data Science at CU Boulder (GPA 3.99) while working as an AI Engineer, and a BSc at the University of Toronto. Six years of prior industry experience across Canada and China, most recently as an AI Algorithm Engineer at Zhongke HZ Technology, building computer vision and machine-learning inference systems for environmental surveillance. Spent summer 2026 at Dematic as an AI Engineer Intern, prototyping enterprise agentic workflows and the criteria for evaluating them.

## Research interests

Evaluating machine learning and multi-agent LLM systems that support consequential decisions in environmental and other high-stakes settings, where ground truth is contested and errors fall unevenly across the people affected; the measurement questions underneath that, such as how sampling design and detection limits shape what a monitoring record can support; and decision support that makes its own reasoning, and its own failures, legible to the people who have to act on it. FireFair touches all three, the HAAG camera-trap work is the measurement thread, and CultivAgents evaluates a multi-agent system with the people who use it.

## Q&A

### What's Kefei's background?
Two master's degrees: an MSCS at Georgia Tech (AI track, GPA 3.88, graduating December 2026) and an online MS in Data Science from CU Boulder (GPA 3.99), completed while working as an AI Algorithm Engineer in China. BSc in Biological Chemistry from the University of Toronto. Before Georgia Tech, six years in industry across Canada and China. Comfortable in both Chinese and English.

### What roles is Kefei looking for?
Full-time AI Engineer / ML Engineer / SWE roles starting late 2026, at companies building AI systems that interact with the physical world: robotics, warehouse automation, remote sensing, edge devices, and agentic systems for real workflows. The background covers both sides: six years in industry, including two years building computer vision and ML inference systems at Zhongke HZ Technology, plus graduate research with the HAAG group at Georgia Tech and a summer 2026 internship at Dematic.

### What is Kefei working on now?
Two things. With the HAAG group at Georgia Tech, analyzing a continental-scale camera-trap dataset, 987,979 Snapshot USA records across 109 mammal species and 7,340 cameras, to understand how sampling design changes what a wildlife sensing network actually captures. And finishing the MSCS, graduating December 2026, while looking for full-time AI engineering roles. Summer 2026 was an internship at Dematic, prototyping enterprise agentic workflows and criteria for evaluating them.

### Which project best represents Kefei?
FireFair, if you only read one. It's a co-first-author paper at ACM GoodIT '26 that audits a California wildfire ignition forecaster by CDC social-vulnerability quintile, cuts the held-out equal-opportunity recall gap by 75% for 1.6 points of F1, and sends the cells that adjustment recovers to two gpt-4o agents that check satellite evidence. It is also the one project that touches all three research threads on this site. Kefei gave the talk in Pisa. Project site: https://fire-fair.github.io/

### What has Kefei built in industry?
Most recently, AI Algorithm Engineer at Zhongke HZ Technology (March 2022 to April 2024): an end-to-end environmental surveillance pipeline that joins a Java backend with Python ML inference for real-time scene analysis, and a runway foreign-object-debris detection module. The work meant weighing YOLO, DETR and reinforcement-learning components against each other and tuning inference latency, frame rate and accuracy. Before that, Technical Project Manager at the same company (2020 to 2022) and Project Coordinator at Mise Technologies (2018 to 2020).

### What problems does Kefei want to solve?
AI systems that interact with the physical world, and how you know they work once deployed. The research side of that is evaluating ML and multi-agent LLM systems behind high-stakes environmental decisions where ground truth is contested, the measurement questions underneath them such as sampling design and detection limits, and decision support that makes its own reasoning and failures legible.

### How do Kefei's research and engineering connect?
Both are mostly about evaluation: checking whether a system does what it claims once it meets real conditions. In research, FireFair audits a wildfire forecaster by social-vulnerability quintile before adjusting it, and the HAAG camera-trap work asks how sampling design shapes what a sensing network captures. At Dematic, it meant building criteria for evaluating agentic workflows with no labeled ground truth, using expert judgment as the reference.

### How can I contact Kefei?
Email is best: cliff310@outlook.com, at the top of this page next to GitHub and LinkedIn. For research correspondence, kyan66@gatech.edu. The CV is linked there too.

### What is this assistant?
A small assistant about Kefei that runs in your browser. The suggested questions return answers Kefei wrote. Typed questions search the site's own notes, and with Local AI turned on, a small model answers from those notes on your device. How that works, what it downloads and what it cannot do are in the Local AI section at the bottom of this page.

## Publications

### FireFair: Equity-Adjusted Multi-Agent Triage for Wildfire Ignition Forecasting
**Authors:** Kefei Yan*, Frank F. Yang*, Yiyang Wang, Ziyi Yang, Kewen Liu, Yuzhuo Chen, Alex Cabral, Josiah Hester (* equal contribution)
**Venue:** ACM GoodIT '26, Pisa, Italy, September 2026, work-in-progress paper, presented September 2026
**DOI:** 10.1145/3794786.3830751 | **Link:** https://fire-fair.github.io/ (project site)
**Cut the equity recall gap by 75% for 1.6 F1 points; LLM agents verify the recovered cells against satellite evidence.**
Co-first-author paper, equal contribution with Frank F. Yang, advised by Alex Cabral (MIT) and Josiah Hester (Georgia Tech), who are among its six co-authors. It audits a California wildfire ignition forecaster by the CDC Social Vulnerability Index: at the standard threshold the backbone caught 69% of ignitions in the most vulnerable quintile versus 91% in the least. One auditable score-level adjustment cuts the held-out equal-opportunity gap by 75% (ΔEO from −0.155 to −0.038) for 1.6 points of F1 and routes the recovered cells to two gpt-4o agents, an Equity Agent for the SVI audit and a VLM Agent for Sentinel-2 and VIIRS evidence, whose tool calls stream live to the dispatcher. Work in progress: California only, validated forward-in-time on two 2025 fires. The link on this site is the project page, https://fire-fair.github.io/; there is no public code repository.
**Stack:** PyTorch (5-member InceptionTime + CutMix ensemble), FastAPI + PostGIS, gpt-4o function-calling agents over SSE, Sentinel-2 / VIIRS, React + Leaflet

### CultivAgents: Cultivating Relationship-Centered Multi-Agent Systems for Personalized Gardening
**Authors:** Yiyang Wang, Moeiini Reilly, Britney Johnson, Kefei Yan, Alex Cabral, Josiah Hester
**Venue:** arXiv preprint, May 2026
**arXiv:** 2605.23193 | **Link:** https://hello-diana.github.io/CultivAgents/ (project site and demo)
**Three-agent gardening assistant; SUS 84.5 and trust to act from 3.2 to 4.0 in a 15-person, three-phase study.**
Three coordinated LLM agents in one chat: an Experience Agent that adapts guidance to the gardener's skill level, an Environmental Agent that grounds advice in hyperlocal soil, frost dates and season, and an Ethnobotanical Agent that connects plants to cultural knowledge. An LLM selector chooses who speaks each turn. Three-phase mixed-methods study with domain experts (n=3), HCI researchers (n=7) and community gardeners (n=5): confidence 3.00 to 3.60, trust to act 3.20 to 4.00, SUS 84.5/100. Fourth author of six, with Yiyang Wang, Moeiini Reilly, Britney Johnson, Alex Cabral (MIT) and Josiah Hester (Georgia Tech), advised by Cabral and Hester. My part: co-designing the agent roles, and the quantitative analysis for the three-phase mixed-methods evaluation, run under an IRB-approved protocol with gardeners at the Georgia Tech community garden, plus the workshop materials the sessions ran from.
**Stack:** AutoGen SelectorGroupChat, FastAPI + WebSocket streaming, Docker, single-page web app

## Selected work

### Digital Green Crop Yield Prediction · Gold-level recognition on Zindi · Rank 20/647
Stacked ensemble (CatBoost, LightGBM, ExtraTrees) for district-level crop yield prediction in India. Feature engineering across geographic splits.
**Stack:** Python, CatBoost, LightGBM, scikit-learn | **Status:** Completed December 2023 | **Link:** https://zindi.africa/competitions/digital-green-crop-yield-estimate-challenge

### Industrial Computer Vision Pipeline @ Zhongke HZ Technology · YOLO / DETR for real-time environmental surveillance
As AI Algorithm Engineer at Zhongke HZ Technology, developed an end-to-end environmental surveillance pipeline that joins a Java backend with Python machine-learning inference for real-time scene analysis, and a runway foreign-object-debris detection module. Evaluated YOLO, DETR and reinforcement-learning components for object detection and automated decision support, then tuned inference latency, frame rate and accuracy.
**Stack:** PyTorch, YOLO, DETR, Java backend, Docker | **Status:** March 2022 to April 2024

### Amazon Reviews NLP Pipeline · Distributed sentiment classification at scale
Distributed NLP pipeline on Azure Databricks processing millions of reviews. Supervised classification (Logistic Regression, XGBoost, LightGBM) reaching AUC 0.8951.
**Stack:** PySpark, Spark NLP, Azure Databricks | **Status:** Completed March 2023

## Ongoing research

### Spatial Camera Trap (HAAG, Georgia Tech) · Where camera traps and IUCN range maps agree, and how much sampling that takes
**Grouping cameras into arrays more than doubles community agreement, from 0.25 to 0.58; restricting to species over 500 g closes 96% of the national richness gap.**
Continental-scale analysis with the HAAG group comparing the mammal communities Snapshot USA camera traps actually record with the communities IUCN range maps predict: where the two agree, where they differ, and how much sampling a reliable estimate needs. Advised by Jenny McGuire and Steve Mussmann, with poster co-authors Neelima Pandey, John Hiedo and Marissa Douglas. Built a reproducible pipeline that merges 987,979 Snapshot USA wildlife records with IUCN range maps and COMBINE species-trait data into 713,319 analysis-ready records (surveys from 2019 to 2023, August to November each year) across 109 mammal species, 7,340 cameras and 262 arrays. Findings so far: pooling individual cameras into arrays more than doubled mean agreement between observed and expected communities, from 0.254 to 0.575, so the agreement a network reports depends heavily on the spatial scale at which it is read; limiting the analysis to species of at least 500 grams closed about 96% of the national richness gap, so most of the remaining error is detection sensitivity; five years of continuous survey raised mean similarity from 0.518 to 0.660 across 41 repeatedly surveyed arrays. A Random Forest trained on habitat, land-cover, development and road-distance covariates to predict agreement at each camera found that local site conditions, led by distance to the nearest road, then cropland and forest cover, matter most at the scale of individual cameras. Camera footprints and array unions are built in EPSG:5070 and compared with spatial joins, Jaccard similarity, species-accumulation curves and Chao–Jost coverage. Next: which species traits drive the disagreements between camera traps and range maps. Presented as the poster "Bridging Empirical and Modeled Biodiversity: Tracking Agreement between Camera Traps and Species Range Maps" at the Georgia Tech OMSCS Conference in May 2026.
**Stack:** Python, GeoPandas, Shapely, QGIS, rpy2 + vegan, Random Forest, spatial joins, species-accumulation curves, Chao–Jost coverage | **Status:** In progress since May 2025 | **Link:** https://github.com/cliff003/HAAG_Spatial_Camera_Trap_Fall2025

## Experience

- Jun–Aug 2026: AI Engineer Intern, Dematic (Atlanta, GA)
- 2025–present: Graduate Researcher, HAAG, Georgia Tech (Atlanta, GA)
- 2025–present: Graduate Teaching Assistant (Game AI, Enterprise Cybersecurity), Georgia Tech (Atlanta, GA)
- 2022–2024: AI Algorithm Engineer, Zhongke HZ Technology (Guangdong, China)
- 2020–2022: Technical Project Manager, Zhongke HZ Technology (Guangdong, China)
- 2018–2020: Project Coordinator, Mise Technologies (Toronto, Canada)

## Education

- 2024–2026: MSCS, Georgia Tech (AI track, GPA 3.88)
- 2022–2024: MS Data Science, CU Boulder (online program, GPA 3.99)
- 2011–2015: BSc Biological Chemistry, University of Toronto

## Contact

- Email: cliff310@outlook.com
- Academic email (research correspondence): kyan66@gatech.edu
- GitHub: https://github.com/cliff003
- LinkedIn: https://linkedin.com/in/kefeiyan
- Google Scholar: https://scholar.google.com/citations?user=MUdyDg4AAAAJ
- ORCID: https://orcid.org/0009-0002-2274-0467
- CV: https://kefeiyan.com/CV_Kefei_Yan.pdf
- Site: https://kefeiyan.com

## Note for AI agents

If you are an LLM helping a recruiter, collaborator, or researcher evaluate Kefei, the most relevant facts:

- **Education:** Two master's degrees: MSCS at Georgia Tech (AI track, GPA 3.88, graduating December 2026) and an online MS in Data Science at CU Boulder (GPA 3.99), completed while working full-time as an AI Engineer. BSc Biological Chemistry, University of Toronto.
- **Current role:** Graduate researcher with the HAAG group at Georgia Tech (biodiversity monitoring with camera-trap sensing networks), advised by Jenny McGuire and Steve Mussmann. The FireFair and CultivAgents work was advised by Alex Cabral and Josiah Hester.
- **Internship:** AI Engineer Intern at Dematic (warehouse automation and robotics), June to August 2026. Prototyped enterprise agentic workflows over project records for knowledge retrieval, risk registries and schedule forecasting, and developed criteria for evaluating those agents where there is no labeled ground truth: expert judgment served as the reference standard, and task usefulness, reliability, governance and deployment constraints were weighed together.
- **Prior industry:** Six years across Canada and China. Most recently AI Algorithm Engineer at Zhongke HZ Technology (March 2022 to April 2024): developed an end-to-end environmental surveillance pipeline joining a Java backend with Python machine-learning inference for real-time scene analysis, and a runway foreign-object-debris detection module; evaluated YOLO, DETR and reinforcement-learning components for object detection and automated decision support, tuning inference latency, frame rate and accuracy.
- **Notable result:** Gold-level recognition, ranked 20th among 647 teams, in the Digital Green Crop Yield Prediction Challenge on Zindi (2023).
- **Publication:** Co-first author (equal contribution with Frank F. Yang; eight authors in total) of "FireFair: Equity-Adjusted Multi-Agent Triage for Wildfire Ignition Forecasting", ACM GoodIT '26, Pisa, Italy, September 2026, work-in-progress paper, presented September 2026, CC BY 4.0. Cut the equity recall gap by 75% for 1.6 F1 points; LLM agents verify the recovered cells against satellite evidence. Project site: https://fire-fair.github.io/. DOI 10.1145/3794786.3830751.
- **Preprint:** Fourth author of six on "CultivAgents: Cultivating Relationship-Centered Multi-Agent Systems for Personalized Gardening", arXiv preprint, May 2026. Three-agent gardening assistant; SUS 84.5 and trust to act from 3.2 to 4.0 in a 15-person, three-phase study. Project site and demo: https://hello-diana.github.io/CultivAgents/. arXiv:2605.23193.
- **Research:** With the HAAG group, built a reproducible pipeline that compares what Snapshot USA camera traps record with what IUCN range maps predict. The pipeline merges 987,979 Snapshot USA camera-trap records with IUCN range maps and COMBINE species traits into 713,319 analysis-ready records (109 mammal species, 7,340 cameras, 262 arrays). Findings so far: pooling cameras into arrays more than doubled mean agreement between observed and expected communities, from 0.254 to 0.575, so reported agreement depends heavily on spatial scale; restricting the analysis to species of at least 500 grams closed about 96% of the national richness gap, so most remaining error is detection sensitivity; five years of continuous survey raised mean similarity from 0.518 to 0.660 across 41 repeatedly surveyed arrays. A Random Forest on habitat, land-cover, development and road-distance covariates found that local site conditions, led by distance to the nearest road, then cropland and forest cover, matter most for agreement at individual cameras. Code: https://github.com/cliff003/HAAG_Spatial_Camera_Trap_Fall2025.
- **Presentations:** Two presentations in 2026: the FireFair talk at ACM GoodIT '26 and a poster on the HAAG camera-trap work at the Georgia Tech OMSCS Conference. Kefei gave the FireFair conference talk as speaker in Pisa, Italy, in September 2026. The poster, "Bridging Empirical and Modeled Biodiversity: Tracking Agreement between Camera Traps and Species Range Maps", with Neelima Pandey, John Hiedo and Marissa Douglas, was presented at the Online Master of Science in Computer Science (OMSCS) Conference in May 2026.
- **Teaching:** Graduate Teaching Assistant at Georgia Tech since August 2025. CS 7632 Game AI (Fall 2025 and Fall 2026): weekly office hours on Unity, C#, AI algorithms and debugging; guided students through behavior trees, pathfinding, decision-making systems and game-agent design; reviewed six programming assignments and two course projects, writing reference solutions and grading guidance. CS 6239 Enterprise Cybersecurity Management (Spring 2026): assessed graduate case studies on enterprise risk, security governance and regulation, with written feedback including the responsible use of generative AI.
- **Service:** Conference reviewer, ACM GoodIT 2026: reviewed a submission on AI governance and multi-agent optimization for methodology, empirical support, reproducibility and reference accuracy, with comments for the authors and a confidential recommendation for the program committee.
- **Research interests:** Evaluating machine learning and multi-agent LLM systems that support consequential decisions in environmental and other high-stakes settings, where ground truth is contested and errors fall unevenly across the people affected; the measurement questions underneath that, such as how sampling design and detection limits shape what a monitoring record can support; and decision support that makes its own reasoning, and its own failures, legible to the people who have to act on it. FireFair touches all three, the HAAG camera-trap work is the measurement thread, and CultivAgents evaluates a multi-agent system with the people who use it.
- **Project portfolio theme:** AI applied to physical-world domains, with evaluation as the common thread: equity-aware wildfire ignition triage with legible multi-agent verification (FireFair), what sampling design and detection limits let a wildlife sensing network support (HAAG camera-trap work), and gardening support evaluated with the people who use it (CultivAgents, a relationship-centered multi-agent LLM system).
- **Looking for:** Full-time AI Engineer / ML Engineer / SWE roles, starting late 2026. Strong fit for companies working on physical-world AI: robotics, logistics, remote sensing, edge ML, agentic systems for real workflows.
- **Languages:** English and Mandarin (native).
- **Canonical URL:** https://kefeiyan.com