Lily Gniedziejko
B.S. Computer Science, University of Illinois Urbana-Champaign
I am an undergraduate researcher in xLab at UIUC, advised by Prof. Tianyin Xu. I work on reinforcement learning environments for evaluating and improving AI agents that diagnose and mitigate failures in production systems.
lilyg3 at illinois.edu · CV · Google Scholar · GitHub · LinkedIn
Education
-
University of Illinois Urbana-Champaign - B.S. Computer Science, expected May 2028
GPA 3.98/4.00 · Dean's List (4×) · James Scholar
-
Highland Park High School - graduated 2024
Salutatorian (2nd of 570) · 4.66 GPA
Research Interests
AI agents are increasingly trusted to diagnose and mitigate failures in critical infrastructure, but how well they actually do so is largely unmeasured. Evaluating these “AI SRE” agents requires environments that reproduce the way real systems break.
I approach this through fault injection at different layers of the stack. In SREGym, I inject faults into live Kubernetes clusters and the distributed systems running atop it. My current work focuses on the network layer: evaluating how well agents perform root cause analysis in emulated data center and campus network fabrics. In addition, building those environments surfaces a prerequisite problem: to check an agent against the intended state of a system, that state has to be specifiable in the first place. My current work also explores how to leverage formal verification to specify networks and distributed system states.
Publications
Experience
- Enabled data centers as a reinforcement learning environment by designing and building a network fault-injection platform for scalable VXLAN/EVPN data center fabrics.
- Built virtualized data centers using VMs running vendor network operating systems.
- Created a live topology explorer that auto-discovers all nodes and links into an interactive graph, letting users click any node to run vendor CLI commands and view output inline.
- Built Grafana dashboards backed by a Prometheus telemetry pipeline for live fabric monitoring.
- Designed a reflective prompt optimization loop that boosted a lightweight model's root-cause-analysis performance to match a frontier model.
- Evaluated AI SRE agents by simulating Kubernetes and multi-cloud infrastructure faults.
- Designed misoperation as a new fault mechanism by deploying TiDB on Kubernetes and building a custom microservice on top of it, extending the benchmark's coverage to management-plane failures.
- Created and deployed the Kubernetes-based application used for TiDB fault injection and mitigation.
- Implemented MCP tools for LangGraph agents, including Jaeger and Prometheus observability tools.
- Built an automated distributed testing tool using tmux-based parallel execution to inject faults across remote nodes.
- Created a trace visualization tool converting JSONL agent outputs into readable HTML, streamlining evaluation and debugging for the research team.
- Developed an internal chatbot that processes over 1,500 technical PDFs, including 400+ page manuals and engineering drawings, to assist the maintenance team, reduce downtime, and link directly to exact pages in source documents.
- Created a Microsoft Teams bot that makes SQL queries and outputs Power BI dashboards.
- Implemented a full-stack data entry application for autopour and melting machines.
Research
SREGym is a high-fidelity, interactive benchmark for AI Site Reliability Engineering: 90 SRE problems spanning hardware, OS, misoperation, and application-level faults across Kubernetes, TiDB, MongoDB, and Kafka. It models production complexity through noise injection and diverse failure modes, including metastable and correlated faults; agents diagnose and mitigate using Prometheus, Loki, and Jaeger MCP servers. SREGym is used by researchers at Microsoft Research, Resolve AI, TierZero (acquired by Cognition), the University of Washington, and SRE startups. [site] [code] [press]
Selected contributions:
- Misoperation fault mechanism. Enabled misoperation as a new fault class by porting TiDB and building a new application on top, expanding the benchmark's fault coverage into a previously unexplored category.
- FleetCast: satellite operations simulator. Real-time simulator for satellite telemetry and orbital passes with a station dashboard showing live satellite connections. Backed by TiDB, containerized with Docker and Helm for Kubernetes deployment; used in xLab for fault injection research.
- Prometheus and Jaeger MCP tools. Model Context Protocol tools for the LangGraph agent so it can query live metrics and distributed traces during incident resolution.
- Agent trace visualizer. Converts JSONL agent outputs into readable, timestamped HTML reports, making evaluation faster and more legible for the research team.
- Fault injection TUI. Interactive terminal UI in Go using Bubble Tea for dynamic fault injection and application deployment.
- Distributed end-to-end benchmark runner. Distributes SREGym problems across multiple nodes, handling cluster creation, parallel tmux execution, and log collection.
Poster Sessions & Talks
- May 2026 - SREGym: A Live Training Ground for AI SRE Agents. Demo presentation at the 1st ACM Conference on AI and Agentic Systems (CAIS '26).
- April 2026 - SREGym paper session. ISUR Research Expo, UIUC.
- November 2025 - Autonomous SRE agent research. Trick or Research, presented to undergraduate students at UIUC.
News
- [TALK] May 2026 - Presented SREGym at ACM CAIS.
- [AWARD] April 2026 - Honored at the 2026 Siebel Celebration of Excellence.
- [PAPER] April 2026 - SREGym was accepted at CAIS '26, the 1st ACM Conference on AI and Agentic Systems, a new ACM venue dedicated to AI agents research. [program]
- [PRESS] April 2026 - SREGym was featured in Siebel School of Computing and Data Science news. [article]
- [GRANT] February 2026 - SREGym received a Slingshot grant from the Laude Institute, supporting continued development of the benchmark and its adoption across academia and industry. [announcement]
- [ADOPTION] SREGym is now used by Microsoft Research, Resolve AI, TierZero, and the University of Washington.
- [AWARD] April 2025 - Honored at the 2025 Siebel Celebration of Excellence.
Awards
- Honored at Siebel School of Computing & Data Science Celebration of Excellence (2025, 2026)
- Dunn Family Scholarship, UIUC
- Engineering Visionary Scholarship, UIUC
- James Scholar, Grainger College of Engineering
- Dean's List, 4 semesters
- NVIDIA Bridge Program, Summer 2025 (selective invitation)
- Salutatorian, Highland Park High School, 2024 (2nd of 570)
- Chamber of Commerce Scholarship, Highland Park, 2024
- Polish School Red Ribbon Honor Graduate (highest honor, every year through graduation)
Service & Leadership
- CS STARS- Student Ambassador & Research Scholar, Siebel School, UIUC (July 2025 – present). I represent the school at events, mentor prospective students, and participate in outreach for the Siebel School of Computing and Data Science.
- Girls Who Code- Facilitator (September 2025 – present). I teach K-12 students Python and mentor them through coding projects.
- Phi Sigma Rho (STEM sorority)- Risk Manager & Social Media Director (September 2024 - December 2025). I led risk management training and support recruitment and service events.