stephen.kim // brisbane, au

Stephen Kim

I build

Five years taking machine learning and generative AI from idea to production — evaluation frameworks with Microsoft Research, forecasting that saved a manufacturer A$1.2M, RAG applications used daily by 500+ people, and agentic pipelines that create video content end-to-end.

SYS.STATUS ONLINE PAPERS 2 SAVED A$1.2M UPTIME 99.9%
Get in touch GitHub ↗ Teaching Portfolio ↗
⌖ PORTRAIT FEED OFFLINE
— set CONFIG.profilePhoto —
FIG. 01 — LIVE · systems designed & shipped to production

AgentEval

GenAI agent evaluation
w/ Microsoft Research · +20% vs G-Eval

TFT Forecaster

Steel-price prediction · GVTL
~A$1.2M saved over 5 yrs

RAG Text-to-SQL

LangChain · Azure
−40% query time · 500+ users

Video Pipeline

Agentic AI · LLM scripts,
voice synthesis, animation

PRODUCTION

Monitored · evaluated ·
used in the real world
FIG. 02 — SELECTED IMPACT

Numbers that held up in the real world

A$0.0M
Cost savings · GVTL

Inventory pipeline + Temporal Fusion Transformer steel-price forecasting for a Vietnamese manufacturer, across five years of live operation.

+0%
vs Google G-Eval

AgentEval, co-developed with Microsoft Research at Telstra — a production evaluation framework for non-deterministic AI agents.

−0%
Query time · 500+ users

Production RAG Text-to-SQL on Azure with LangChain, prompt engineering and ADK agent orchestration.

A$0K
First Prize · CSIRO

Won a CSIRO innovation program with an AI-powered app supporting plant-based meat and sustainable consumption.

0
First-author papers

Peer-reviewed publications on AI evaluation and AI-driven productivity (Springer CCIS; CS&IT).

FIG. 03 — LIVE DATA CONSOLE

Don't take my word for it — touch the data

Delivered improvements
measured outcomes across shipped systems (%)
Capability radar
self-assessed — happy to be challenged in an interview
Forecast playground
illustrative demo — the production version was a Temporal Fusion Transformer forecasting steel prices (GVTL, ~A$1.2M saved)
horizon 6 mo
GitHub telemetry
live from api.github.com — refreshes on every visit
—
public repos
—
followers
—
days since last push
    FIG. 04 — MISSION TIMELINE

    Nine missions, four parallel tracks

    01 / 09
    NOW
    consulting

    Open to the next mission

    AI Engineer / Data Scientist roles · Brisbane (hybrid or onsite)
    immediately available
    • Production GenAI & ML, end-to-end data engineering, and a habit of leaving systems better than I found them
    ▸ thanh31596000@gmail.com
    2024
    industry

    Data & AI Specialist

    Telstra Limited Group · Brisbane
    MAR — SEP 2024
    • Co-developed AgentEval with Microsoft Research — +20% over Google's G-Eval
    • Production RAG Text-to-SQL (LangChain, ADK) on Azure — −40% query time, 500+ users
    • PGI data-quality framework (35% accuracy gains); co-authored 2 papers
    2022–26
    consulting

    ML Engineer / Data Consultant

    GVTL · Vietnam (remote) · 4+ years
    2022 — 2026
    • Owned core data assets end-to-end: inventory ETL pipeline in daily operation
    • Temporal Fusion Transformer (PyTorch) steel-price forecasting — ~A$1.2M saved
    • Full lifecycle: ingestion → features → training → deployment → monitoring
    2026
    research

    PhD — thesis lodged

    QUT · Prof. Richi Nayak · RPA Stipend Scholarship
    2022 — 2026
    • From Data Enrichment to Responsible Recommendations
    • Five PyTorch architectures incl. an agentic LLM with hallucination mitigation
    2022–
    teaching

    Casual Lecturer · Subject Coordinator

    Kaplan · QUT · JCU · CQU
    JUL 2022 — ONGOING
    • 13+ units: ML, deep learning, NLP, statistics, data communication
    • DATA5000 curriculum, assessments & rubrics; 2024 QUT Teaching Advantage Program
    2023
    industry

    Machine Learning Engineer

    Prooftec · Sydney (remote)
    JAN — MAY 2023
    • Scalable production ETL pipelines (dbt on AWS); automated testing & lineage
    • Docker + CI/CD; automated retraining; 10% infra cost cut
    2021–22
    industry

    Data Scientist

    Orefox Limited Co · Brisbane
    OCT 2021 — OCT 2022
    • Predictive models over geospatial data (PostGIS) — +20% site identification
    • NLP information-extraction engine — 30% faster retrieval
    2021–22
    industry

    Junior Software Developer

    Explorate · Brisbane
    NOV 2021 — SEP 2022
    • Production REST APIs on AWS — 10K+ daily requests, 99.9% uptime
    • Document OCR automation with Textract
    2021–22
    research

    Research Assistant ×2

    QUT · Brisbane
    NOV 2021 — AUG 2022
    • RL in recommender systems; deep matrix factorisation for text classification
    FIG. 05 — CAPABILITIES

    Tools of the trade

    AI / GenAI

    LLMsRAGAgentic AI LangChainPydanticADK Prompt engineeringAI evaluation Hallucination mitigationResponsible AI

    Machine Learning

    PyTorchTensorFlowScikit-learn TFT / time-seriesGradient boosting NLPDeep learningRecommender systems

    Engineering & Cloud

    PythonSQLRC/C++ AzureAWSDocker CI/CDMLflowdbtPySpark

    Data & Communication

    PostgreSQLMongoDBPower BI TableauDimensional modelling Teaching & storytelling
    FIG. 06 — PERSONAL PROJECT

    The machine that makes the videos

    Founder & Creator · 2025 — present

    AI-Powered YouTube Content Channel

    An agentic AI pipeline that produces video content end-to-end — no camera crew, no editor, no studio. One orchestrated system, from idea to upload.

    • LLM script generation — story and narration written by language models with structured prompting.
    • Voice synthesis — natural narration generated programmatically.
    • Automated animation — visuals rendered by code and composited automatically.
    • Tool orchestration — every stage wired together in Python as a single production pipeline.
    NO SIGNAL PROJECT PHOTO FEED
    — set CONFIG.projectPhoto —
    NO SIGNAL VIDEO FEED
    — set CONFIG.youtubeVideoId —
    FIG. 07 — PUBLICATIONS (FIRST AUTHOR)

    On the record

    SPRINGER CCIS VOL. 2325 · AUSDM 2024 · DOI: 10.1007/978-981-95-6888-8_1

    Generative Agents as Reliable Proxies for Human Evaluation of AI-Generated Content

    Vu, T., Nayak, R., & Balasubramaniam, T. (2026) — can AI agents be trusted to judge AI output? A framework and evidence.

    CS & IT VOL. 14, NO. 24 · DOI: 10.5121/csit.2024.142402

    Improved Productivity with AI Models for SQL Tasks: A Case Study

    Vu, T., Keretna, S., Nayak, R., & Balasubramaniam, T. (2024) — measuring real productivity gains when AI assists enterprise SQL work.

    FIG. 08 — EDUCATION & RECOGNITION

    Foundations

    APR 2022 — 2026 (LODGED)

    PhD, Computer Science — QUT

    QUT RPA Stipend Scholarship (competitive merit-based). Thesis on responsible AI and recommender systems — five PyTorch architectures including an agentic LLM with hallucination mitigation. Supervisor: Prof. Richi Nayak.

    FEB 2020 — OCT 2021

    Master of Information Technology — QUT

    GPA 6.5 / 7.0 (High Distinction) · Dean's List 2020 & 2021.

    ★

    First Prize & A$20,000 — CSIRO innovation program, AI app for plant-based meat & sustainable consumption.

    ★

    2024 QUT Teaching Advantage Program — professional teaching certification.