FIG.0 — SURFACESHEET 1 / 4

Thiraput Khongmuak

ML engineer. Every layer of the stack.

EMAILGITHUBLINKEDINMEDIUMRÉSUMÉ (PDF)
FIELD RECORD — TMBTHANACHART BANK฿6M refinance volume from customer-data analysis, presented to non-technical stakeholders
Thiraput Khongmuak
FIG.0 — THE ENGINEER
DEPTH 0 / 3SCROLL TO DESCEND

FIG.1 — SYSTEMS · HOW IT RUNS

Two systems, built end to end.

Drills.pro

01 — PROBLEM

Training and deploying a model takes a team — data pipeline, GPUs, serving, monitoring. Most ideas die before any of that exists.

02 — BUILT

A low-code platform that takes a dataset to a deployed model: Go microservices behind one gateway, async training over RabbitMQ, NVIDIA Triton serving with TensorRT, Loki and Grafana watching all of it.

03 — RESULT

One person can now do the team's job — and the serving layer runs faster than the baseline it replaced.

+43%

inference throughput

55.5%

training speedup

0→1

platform, solo-built

→ drills.pro
Drills.pro screenshot
FIG.1A — DRILLS.PRO · TRAINING DASHBOARD

Traffy-Caesar

01 — PROBLEM

Bangkok's Traffy Fondue collects thousands of citizen complaints a day. Routing them is manual and slow.

02 — BUILT

A daily Airflow + PySpark ETL that embeds every complaint — text and image — clusters them with BigQuery ML, and hands operators a dashboard with Gemini-written summaries per cluster.

03 — RESULT

City-scale complaints, grouped and explained before a human reads them.

daily

city-scale ETL

text+image

embedded modalities

→ www.youtube.com
Traffy-Caesar screenshot
FIG.1B — TRAFFY-CAESAR · CLUSTER DASHBOARD
YES CAMP · MITR PHOL — WINNERAI knowledge hub on AWS — ฿3.1M/month productivity impact
SCBX × AIAT — 2ND RUNNER-UPself-correcting multi-agent financial analysis on LangGraph

FIG.2 — ARCHITECTURE · HOW IT WORKS

Reading the source of the field.

INSTRUMENT 01 — RUN IT YOURSELF

single-head scaled dot-product attention — softmax(QKᵀ/√d) — implemented from scratch in TypeScript. toy projections, real math. hover a token to see where it attends.

NECTEC — RESEARCH ASSISTANTThai voice agent on Gemini Live — tool calling, memory, robotics

FIG.3 — FIRST PRINCIPLES · WHY IT WORKS

If I can't derive it, I don't claim it.

THE INDEX — EVERYTHING ELSE

yrworktag
2026PeaceKeeperllm
2025CP-GENIE rag
2025Super AI Engineer S5award
2025I-Squared Hackathonaward
2025Traffy Multilabelnlp
2025PicMePls web
2025Backend TA — Chulalongkornrole
2024Chess → PGN vision
2024Courspora Notifier infra
2024TEEYAI web
2025Writing — four deep dives writing