Open to AI/ML · Data · Agent-systems roles — Taipei / remote

Portrait of Kuan-Hao Tseng

Kuan-Hao Tseng曾冠豪

I build LLM agents that troubleshoot real computer networks.

MSc Computer Science (research pathway) at the University of Sydney, graduating July 2026. First author of SADE — accepted at IEEE LCN 2026 — and upstream contributor to the NIKA network-diagnostics benchmark.

Read the paper GitHub LinkedIn Email
sade — diagnostic session
$ sade diagnose --incident nika-0231
✔ symptom intake    intermittent loss reported on edge fabric
▸ phase gate       evidence first — no hypothesis committed yet
✔ layer escalation  L1 ok · L2 clean · L3: OSPF stuck in EXSTART
✔ targeted probe    show ip ospf interface → MTU 1500 ≠ 9000
✔ hypothesis commit root cause: MTU mismatch on ge-0/0/1
— localized · NIKA benchmark · root-cause F1 +37pp vs ReAct baseline

Illustrative session — SADE encodes the layered troubleshooting discipline of human network engineers as an explicit, phase-gated agent policy.

Research

IEEE LCN 2026 · accepted first author arXiv:2605.04530

SADE: Symptom-Aware Diagnostic Escalation for LLM-Based Network Troubleshooting

Kuan-Hao Tseng, Niruth Bogahawatta, Yasod Ginige, Kosta Dekic, Arunan Sivanathan, Suranga Seneviratne

LLM agents applied to network troubleshooting tend to conflate gathering evidence with committing to a hypothesis. SADE encodes the classical layer-by-layer troubleshooting methodology as an explicit policy: a phase-gated diagnostic workflow that separates evidence acquisition from hypothesis commitment, paired with a routed library of fault-family skills and high-yield diagnostic helpers.

+37pp
root-cause F1 over a ReAct + GPT-5 baseline
22pp
attributable to the diagnostic policy alone, same-model comparison
523
held-out NIKA incidents across 11 unseen scenarios
arXiv preprint SADE-NetworkAgent →

Experience

University of Sydney — NSS Lab Student Researcher Jul 2025 — present
  • Built network-research datasets and simulation environments across Cisco Packet Tracer, GNS3, and Kathara; migrated the lab's emulation stack to Docker-based Kathara for reproducible, deployable experiments.
  • Designed a Configuration → LLM → Topology static-analysis pipeline: parse device configs with an LLM and reconstruct L3 topology into Neo4j for Graph-RAG queries and relationship analysis.
  • Built the dynamic-analysis line — MCP servers and the Claude Agent SDK driving live emulated networks on the NIKA benchmark — the work that grew into SADE.
Granite River Labs (GRL) FV Test Engineer Intern · Taipei Sep 2023 — Jan 2024
  • Ran hardware-validation projects on C and Arduino test programs; tuned existing test suites for throughput and result accuracy.
  • Owned test execution, progress tracking, and issue reporting on customer projects.

Projects

SADE-NetworkAgentfeatured

Phase-gated LLM diagnostic agent for root-cause localization in computer networks. The reference implementation behind the IEEE LCN 2026 paper.

PythonClaude Agent SDKNIKA
root-cause F1 +37pp vs ReAct baseline
nikaupstream

Contributor to sands-lab/nika — the public benchmark of reproducible network-fault incidents on emulated topologies that SADE is evaluated on.

KatharaDockerbenchmarks
contributions merged upstream
au-energy-data-pipeline

ETL over Australian public energy data — NGER, CER, and ABS sources scraped, normalized, and geocoded into a DuckDB warehouse with spatial layers.

PythonDuckDBETL
4.2M rows, end-to-end reproducible
nem-power-data-pipeline

Streaming pipeline for Australia's electricity market: REST acquisition, DuckDB star schema cross-linked to the CER register, MQTT replay, and a live Streamlit + Folium map dashboard.

DuckDBMQTTStreamlit
716,454 obs · 356 facilities · 5 s live repaint
used-car-price-prediction

Price-regression benchmark on 14,699 cleaned listings: linear regression, decision tree, and random forest with full regression diagnostics and feature-importance analysis.

scikit-learnpandas
random forest R² 0.911 · MAE ≈ A$1.9k
cnn-image-classifier

CIFAR-10 three-way benchmark — Random Forest vs MLP vs CNN on identical splits, showing exactly what convolutional structure buys on images.

TensorFlow / Kerasscikit-learn
CNN 84.2% test acc vs MLP 50.7% · RF 46.3%
airline-review-satisfaction

Three-class satisfaction model from ~22k Skytrax airline reviews — five classifiers benchmarked with and without SMOTE, leak-free imbalanced-learn pipelines throughout.

scikit-learnXGBoostSMOTE
RF + SMOTE macro F1 0.693 · acc 82.3%
car-dealership-db

Relational design for a multi-site dealership group: normalized PostgreSQL schema, ERD documentation, and a set of analytical queries.

PostgreSQLSQL
normalized schema + ERD + query suite

Skills

Languages
PythonSQLRCC++C#
ML / DL
PyTorchTensorFlow / Kerasscikit-learnpandasNumPy
LLM & agents
Claude Agent SDKMCPRAG / Graph RAGLangChainChromaDB
Data & infra
PostgreSQLNeo4jDuckDBDockerGit
Networking
Cisco IOS configsGNS3Packet TracerKatharaL2/L3 diagnostics

Education

University of Sydney
MSc Computer Science, research pathway · 2024 — Jul 2026 (expected)

GPA 6.98 / 7.0 (WAM 84.8) — Distinction average. Machine Learning & Data Mining 92 (HD) · Principles of Data Science 88 (HD) · Database Management 82 (D).

National Taiwan University of Science and Technology
BEng Mechanical Engineering · 2019 — 2023

Self-directed pivot from mechanical engineering into software and AI — C# and app development first, then Python, ML, and the web stack.

Human languages: Mandarin (native) · Taiwanese Hokkien (native) · English (professional) · Japanese (basic)

Let's talk.

I'm graduating in July 2026 and looking for AI/ML, data-engineering, or agent-systems roles in Taipei / New Taipei or remote. If my work fits your team, my inbox is open.

jeffqaz741@gmail.com GitHub LinkedIn