KylianPersonal site
Email中文

Work / selected cases

These are the things
I genuinely spent time making.

Across agent traces, information overload, and GIS code, I start by locating what actually blocks the outcome. Then I decide whether the answer is a workflow, an evaluation, or a tool. These three cases document that process.

TB2 Diagnostics
TRACECAUSEFIX

Agent diagnostics · Evaluation system · 2025–26

Agent Failure Diagnostics

TB2 Diagnostics

Turning long agent-failure traces into evidence-backed error types and repair priorities, reducing the cost of manual review.

Why it mattersStructured diagnosis turns “this agent feels unreliable” into “which layer failed, what evidence supports it, and what to fix first,” giving researchers and evaluators a shared review language.

What I worked onI worked on failure analysis, taxonomy, evidence-chain design, the LLM-as-Judge flow, and human–model agreement validation.

Read the full case ↗
FOMO Firewall
LEARNLATERIGNORE

Information triage · Product judgment · 2026

Information Triage & Learning Assistant

FOMO Firewall

An information-triage prototype for knowledge workers that turns endless saving into three executable choices: Learn now, Later, or Ignore.

Why it mattersIf triage works, users save less without acting and reach learning faster. The system can also learn from acceptance and overrides instead of permanently deciding on the user’s behalf.

What I worked onI independently shaped the problem, interaction path, triage logic, layered memory, and validation design.

Read the full case ↗
GeoAgent
SEARCHRUNVERIFY

GIS · RAG · Multi-Agent · 2025

Verifiable GIS Code Assistant

GeoAgent

Turning GIS code generation from a one-shot answer into a loop of document retrieval, generation, execution, repair, and evaluation.

Why it mattersConnecting external knowledge, execution feedback, and evaluation makes errors traceable to a specific stage and lets users judge whether output is runnable, repairable, and deliverable.

What I worked onI focused on the RAG knowledge base, task data, evaluation analysis, and the path from natural language to generated code within a collaborative research project.

Read the full case ↗

Field experience

High-Definition Map Intern

Tianyi Transportation Technology · Suzhou

Researched the boundaries of formats including GeoJSON and OpenDRIVE; delivery focused on SHP/QGIS road-element work, field-change verification, and platform replacement tests. One road change was corrected and validated within three days, followed by a QGIS plugin demo and operating guide.