About Experience Floodtir Hackathons Projects Stack Awards Contact

Super AI Engineer S5 · 🥉 Bronze Medalist

Natthawin
Tamprasert.

AI & ML Engineer Data Scientist IoT Systems Engineer

I build AI systems that ship and make impact — from time-series forecasting and NLP to computer vision and RAG / LLM pipelines. Dual-degree student in Industrial Physics & IoT Engineering @ KMITL.

GitHub View my work

Pathum Thani, TH natwin1313@gmail.com 098-832-5128

🥉

Individual

Bronze Medal

4/7

National
hackathon wins

Recognized by AI Builders AIAT BDI · Big Data Institute CP AXTRA Business Online (BOL) Chula CPALL · 7-Eleven Sony Dhipaya Insurance KMITL

🥉

Bronze Medal — Super AI Engineer S5, overall national ranking (AIAT)

4/7

Podium in national industry hackathons: 3× 1st, 1× runner-up

119K+

Real insurance claims behind an explainable Vision-AI pipeline

Top 21

BDI Hackathon 2026 finalist — flood depth from one photo

about_me.py

class NatthawinTamprasert:
    def __init__(self):
        self.role     = "Data Scientist & AI Developer"
        self.degree   = "Physics + IoT Eng (Dual)"
        self.school   = "KMITL"
        self.medal    = "🥉 Bronze · Super AI S5"
        self.record   = "Won 4 of 7 hackathons"
        self.interests = [
            "Time-Series", "NLP",
            "Computer Vision",
            "RAG / LLM", "IoT + AI",
        ]

    def philosophy(self):
        # The best model is the one that
        # ships and makes impact.
        return self.build_and_deploy()

About me

AI engineer who turns research into shipped systems.

I'm Natthawin Tamprasert (ณัฐวินท์ แต้มประเสริฐ), a dual-degree student in Industrial Physics and IoT Systems Engineering at KMITL, and a Bronze Medalist of the national Super AI Engineer Season 5 program.

My path into AI started at AI Builders (2022–23) with a rain-prediction model. Since then, across seven national hackathons with Team Pangpuriye I won four — spanning demand forecasting, Thai legal NLP, and medical imaging — and interned building an explainable Vision-AI pipeline for vehicle-damage assessment on 119K+ insurance claims.

Most recently I was the model architect of Floodtir, a BDI Hackathon 2026 Top-21 finalist that measures flood depth from a single street photo.

Time-Series Forecasting NLP Computer Vision RAG / LLM Systems IoT + AI Integration

Experience

AI Research Internship

An explainable, end-to-end Vision-AI system built during my internship — designed to make insurance claims fairer and more transparent.

🏢 Vehicle Damage Assessment · 2025

AI Research Intern

Dhipaya Insurance (ทิพยประกันภัย)

Developed an explainable 4-stage Vision-AI pipeline addressing a 17.5% valuation discrepancy across 119,212 insurance claims — pairing deep-learning detectors with an LLM report generator for plain-language, auditable output.

↑ Transparency ↑ Negotiation leverage ↓ Claim costs

Architecture diagram of the 4-stage vehicle damage assessment pipeline

System architecture — click to zoom

Car / Non-Car Filter

CLIP Prompt Ensembling · zero-shot

Damage Localization

DINOv2 + Top-K Attention Pooling · Hamming 0.84

Damage Classification

ResNet-101 + DCN + FPN · 20+ classes

Damage % Estimation

Geometric: A_crop = A_full / Z²

✦

Report Generation

LLaMA 3.1-8B (4-bit) · plain-language output

119K+real claims analysed

17.5%insurer vs. repair-shop gap

20+damage types classified

Latest project · 2026

A flood gauge from any street photo

Team Floodtir at the BDI Hackathon 2026 finals booth

🏁 Finalist · Top 21 teams

Big Data Institute (BDI) Hackathon 2026 · Jun – Jul 2026

Floodtir — Model Architect

Bangkok's 562 water gauges sit on main roads, yet 45% of its roads are dead-end sois with no data at all. With Floodtir, anyone can report a flood by sending a photo on LINE. The model I designed, FAHRPF, reads the water depth from that photo, with no sensors installed.

16.01 cmmean error

−77.8%params vs a 4B VLM

0labelled training images

Zero-shot pipeline of SAM 3 · YOLO26 · Depth Anything V2, where cars and people act as rulers for the water line.
Within 1 cm of a 4B-parameter VLM at 893M parameters, so it can run on-premise and at the edge.
Built the evaluation set from scratch from real Thai flood news imagery.

How FAHRPF works

Water line — SAM 3 segments the water
Reference objects — YOLO26 + SAM 3 instance masks
Depth — Depth Anything V2 (monocular)
Per object — submerged ratio × known object height
Fuse — average weighted by depth × confidence

Why it matters: Thai floods have cost 2,000+ lives and ฿2T+ over 20 years, and Bangkok loses up to ฿300M per hour at peak. Floodtir is built for the Bangkok Governor's office, which oversees the city's flood response.

Hackathons

Podium in 4 of 7 national hackathons

Super AI Engineer Season 5 industry hackathons with Team Pangpuriye: one week per problem, on real company data.

Demand Forecasting Hackathon certificate

View

🥇 1st Place

AIAT × AXONS

Demand Forecasting

May 2025

EDA & solution design

Led EDA revealing category-specific demand; shifted strategy from global → category-wise modeling
Fine-tuned Chronos time-series models with custom post-processing
PCA + KMeans clustering to capture seasonal patterns & variability

View

🥇 1st Place

AIAT × BOL

Crime Charges Analysis · Thai NLP

May 2025

NER model & legal RAG

Chose PhayaThaiBERT (F1 0.985) over the larger XLM-RoBERTa: better on Thai at lower cost
NER name-masking to reduce demographic bias in charge prediction
Legal RAG (FAISS) answers ambiguous cases from statutes, not guesswork

Liver Fibrosis Severity Prediction certificate

View

🥇 1st Place

AIAT × CHULA

Liver Fibrosis Severity Prediction

Jun 2025

Lead model developer

Predicted liver stiffness from ultrasound, a low-cost alternative to specialised hardware
Found tree-based heads beat MLPs on small data, redirecting the architecture
Domain-aware post-processing cut MAE 2.9 → 2.4

View

🥈 1st Runner-up

AIAT × CP AXTRA

Sales Forecasting & Product Optimization

May 2025

Lead EDA

Profiled sparsity and seasonality per SKU to decide ML vs. statistics per product
Batch processing + stratified sampling to fit large retail data in memory

Other challenges

View

🏪 LIFT

CPALL × Creative AI Club

CAI Club Hackathon · 7-Eleven × AI

Jan 2025 · Invited speaker

Built LIFT — social-listening system monitoring Thai news to predict demand spikes at 7-Eleven
News ingestion → LLM summarization → Typhoon API sentiment → high-demand flagging
Estimated ฿42.5M opportunity-cost savings from preventing a viral-product stockout

View

🏭 Top 10

Sony · KMITL

Smart Factory Hackathon 2025

Supply-Chain AI System

End-to-end AI for inventory analysis, risk assessment & automated procurement
GMM unsupervised clustering to surface anomalies without labels
Deployed Qwen 2.5 GGUF fully on-premise — zero data leakage

Projects

Selected work

Open-source projects and research systems — full code and write-ups on GitHub.

Natthawin presenting the rain-prediction poster at AI Builders

AI Builders · 2022–23 Showcase Series #67

Will it rain tomorrow? Where my AI journey started

Predicts tomorrow's rain from 3,249 days × 20 variables of TMD weather-station data. Revisited after Super AI Engineer: 67% → 82% accuracy.

AI Builders · 202367%

Revisited · post Super AI82%

Caught a silent data leak: missing days made the model predict two days ahead. Rebuilt a full calendar with a 1-day lag join and dropped them.
Lagged rainfall, pressure and temperature features, chosen by feature importance and explained by the physics of condensation.

ClassificationTime-SeriesFeature Eng.Weather

Read on Medium View repo

🌾

Jupyter

csiro-image2biomass

Top 10.1% Kaggle solution using DINOv3 + FiLM for pasture biomass prediction from imagery.

DINOv3FiLMCVRegression

View repo 1

🌊

Jupyter

chaophraya-water-quality

IoT-based water-quality forecasting benchmarking XGBoost · CatBoost · LSTM · GRU · Transformer on 10-year time-series.

Time-SeriesIoTBenchmark

View repo 1

🎙️

Jupyter

SpExNote

Intelligent lecture-recording pipeline: Silero-VAD → SpEx+ speaker extraction → Thai ASR → Qwen-RAG summarization.

SpeechThai ASRRAGIoT

View repo

🤖

TypeScript

NongAuanIoT

Thai multi-agent RAG chatbot for KMITL IoTE — Intent Classifier routes to Academic · TCAS · General agents on Typhoon-S 8B, fully CPU-based.

Multi-AgentRAGLLMThai

View repo

💬

Jupyter

Chat-bot-Jar-neng

Thai-language chatbot with synthetic data generation and a fine-tuned LLM + RAG retrieval stack.

LLM Fine-tuneRAGThai

View repo 2

See all repositories on GitHub

Tech stack

Tools I build with

🧠 Data Science & ML

PythonPyTorchTensorFlowscikit-learnPandasNumPyXGBoostLightGBMCatBoostChronos

✨ LLM & Generative AI

Hugging FaceAutoGluonFAISSRAG ArchitectureLLaMA 3.1Qwen 2.5 GGUFTyphoon-SPhayaThaiBERTXLM-RoBERTa

👁️ Computer Vision

SAM 2 / 3YOLO v8 / 26Depth Anything V2DINOv2 / v3CLIPEVA02ResNetFlorence-2OpenCV

🔌 IoT & Hardware

ArduinoNode-REDMQTTNB-IoTSensors

⚙️ DevOps & Tools

FastAPIDockerGoogle CloudFirebaseRedisGitMATLAB

</> Languages

PythonC++CJavaScriptTypeScriptNode.js

Awards & Certificates

Proof of work

Competition certificates sit on each hackathon card above. Below are training programs, credentials, and moments from the journey. Click any image to view full size.