Phattaraphol Unkamnerd

As an AI & Machine Learning Engineer, I build systems that keep running after the demo ends. If a system still needs me to run it, I don't consider it finished. That pushes me to automate the whole path, from synthetic data generation and auto-labeling through training, export, and on-device inference, and to validate what comes out of a model instead of trusting it.

I specialize in Generative AI, RAG, and multimodal pipelines, and I own the systems I build end to end, from architecture to deployment.

AI and Machine Learning Engineer at CUBIC X Co., Ltd., a spin-off of Thaipet Industrial Co., Ltd. (2025-08 – Present)

• Object-Aware Conversational AI Platform: Built a FastAPI microservice backend that lets products answer questions in first person over text, voice, and image. Grounds answers with a dynamic RAG system (Qwen3.5/Gemma3 on Ollama, ChromaDB vector store), and handles input via Whisper for STT, a dual-backend TTS layer (JaiTTS for Thai, KittenTTS for English), and fastText for language detection.
• Automated Product Metadata Pipeline: Built an agentic vision pipeline turning 3D scans and photos into verified, structured metadata with zero manual labeling. Uses Qwen-VL to auto-label class and brand, Grounding DINO for open-vocabulary detection, SAM 2.1 for segmentation, and dual OCR (PaddleOCR, TyphoonOCR), then a Qwen3 LLM merges and verifies it into schema-validated output. Deployed as a Dropbox-driven service.
• Synthetic Data & CV Training Pipeline: Built an end-to-end pipeline generating synthetic datasets from 3D models via Unity Perception (HDRP) with domain randomization, merged with real photos auto-labeled by reused Grounding DINO + SAM 2.1 modules. Automated SOLO to COCO to YOLO conversion and trained YOLO, RT-DETR, and DINOv3 detectors to ~0.7–0.8 mAP, with synthetic data improving recall and keeping detection robust across camera angles. Tracked via Comet ML and TensorBoard and exported to ONNX and Unity Sentis for on-device inference.
• Image-Guided 3D Reconstruction Pipeline: Built a microservice pipeline that repairs photogrammetry scans of transparent/reflective products that often scan with holes on iPhone LiDAR. Used OpenCV to flag specular reflections in iOS Object Capture frames and routed them through a ComfyUI workflow running Qwen-Image-Edit (Lightning LoRA for fewer diffusion steps), parameterized per-node from Python, then merged the cleaned frames back into Apple Object Capture for a cleaner mesh.

Data Scientist Intern at Supalai Public Company Limited (2024-11 – 2024-12)

• Defect Forecasting: Analyzed and forecasted housing defects in low-rise construction projects.
• BOQ Item Classification: Built a Thai-language classifier mapping construction bill-of-quantities line items into cost categories from five tabular fields, benchmarking CNN, RNN, LSTM, and BiLSTM architectures against a GPT API baseline and scoring per-class precision and recall with scikit-learn.
• Data Collection Pipelines: Wrote Python scrapers gathering competitor news and daily commodity and macro series covering metals, oil, FX, and Fed data into datasets feeding internal Tableau dashboards.
• Data Scraping: Scraped construction material prices into Tableau with indicators, and Emission Factor data for Carbon Footprint (CFO) analysis.
• Data Vectorization: Processed and vectorized data to train the company's chat framework.
• Document AI: Built a PDF-to-text converter using Azure Document Intelligence (Python).

Data Analyst Intern at Supalai Public Company Limited (2024-07 – 2024-10)

• AutoML Deployment: Prepared datasets and trained models with AutoGluon, deploying the best via FastAPI.
• Computer Vision: Trained YOLO models for a heatmap generator and person detection with dwell-time tracking.
• Data Lineage: Used MongoDB to reverse-engineer and trace data origins in summary reports.
• Data Pipelines: Built Azure Synapse pipelines feeding a data warehouse.

Bachelor of Science in Data Science – Silpakorn University (2021-07 – 2025-04)