APPLIED AI ENGINEER

AI systems built with
production discipline.

Seven years building production AI, NLP, and data systems. Today I focus on reliable RAG, agents, and evaluation.

View Work

Selected work

Production work and controlled experiments, with the evidence labeled clearly.

Jupiter Career Agent

A live career agent with hybrid retrieval, grounded answers, and evaluation at each layer.

0.919retrieval ranking · offline

0.923generation faithfulness · RAGAS

99.3%intent routing accuracy

LangGraphHybrid retrievalRerankingRAGAS

Try Jupiter ↗

Aunalytics — banking NLP in production

Led a community-bank Text-to-SQL modernization that improved accuracy and cut onboarding time.

~20%SQL accuracy improvement

~4 weeks → 1 sprintclient onboarding

0production hotfixes

Text-to-SQLSchema groundingValidation + retryRegulated finance

FinScope — distilling research behavior into a 4B model

Distilled executed Gemini teacher tool-use trajectories into Qwen3-4B with QLoRA. In a controlled comparison, the base hallucinated sources; the adapted model searched first and used retrieved sources.

Qwen3-4BQLoRABehavior distillationSelf-hostable inference

Recorded controlled comparison Same prompt · adapter off vs. on

Same prompt and tools; only the adapter changes. Base: answers directly with invented sources. Adapted: searches first and brings retrieved sources into the answer.

RetailOps — support agent with tool guardrails

A deployed e-commerce proof of concept with enforced policies, manager approval, and database-state evaluation.

91%tool-selection pass rate

96%tool-efficiency pass rate

Google ADKHuman approvalState-based evals

guided replay· recorded verified paths

Experience

// Background

~4.5 years at Aunalytics, including ~2.5 years leading NLP delivery for community-bank products. Independent Applied AI work since 2024; built and launched Jupiter in February 2026.

M.A. Statistics · ColumbiaECML PKDD 2021 · Springer LNCS

What I build

Evaluation & Reliability

Benchmarks, retrieval metrics, behavioral checks, and error analysis.

Applied AI & NLP Systems

RAG, Text-to-SQL, routing, adaptation, and agent workflows.

Production Delivery

Python services, guardrails, monitoring, and cross-functional delivery.

Technical communication // @LLMImplementation

Working code, clearly explained.

I also publish code-first walkthroughs on RAG, agents, model adaptation, evaluation, and deployment.

View technical walkthroughs ↗

LET'S TALK

Let's build something reliable.

Open to hands-on Applied AI / ML roles and selective product collaborations across Canada.

shane.xia.work@gmail.com LinkedIn ↗