Ati Tesakulsiri

Working Experience

Selected Work Projects

• Text-Based Video Smart Search

  Developing a PoC surveillance-record search engine with Thai-language generative AI. Leveraged LangChain, SmolAgent, and Ollama for on-premise LLM serving; developed vision-language multimodal systems, fine-tuned models, and vector database integration for video analytics programs.

• CCTV Black Smoke Measurement and Evaluation

  Developed research software to detect and analyze black smoke from vehicles in collaboration with ENTEC, NSTDA. Built a web application with video analytics for smoke screening and multi-frame emissions analysis from light-duty diesel trucks and vans.

• Safety AI PoC

  Developing a tollway surveillance PoC to detect wrong-way driving, speeding, illegal parking, and unexpected objects. Built a zoom-assisted API for falling-object detection up to 300 meters and vehicle length classification.

• AI Service for Video Analytics

  Contributed to AI service tasks, AI training, optimization, hosting tests, and CI pipeline upgrades for automated unit tests, evaluation tests, reporting, and service software patching.

• Face Recognition

  Owned the analytics-side face recognition system, including detection, tracking, recognition integration, evaluation, and deployment support.

• Wildfire Simulation and Component Classification

  Developed wildfire component classification and segmentation from drone video, with forecasting support using empirical models and satellite data.

• Video Action Recognition

  Improved action recognition post-processing to reduce fall detection false positives and performed system performance evaluation.

• Other Demo Projects

  Developed accident detection demo logic using tracked object pair information and Qwen 3.5 for reasoning over vehicle interactions. Built VLM-based scene detection for fire, violence, fall, smoke, and abnormal surveillance events.

University & Personal Projects

• Joules of Saim

  Designed and built an end-to-end business intelligence project to forecast electricity prices, including data gathering from BOT and EGAT APIs, time-series model training, and an interactive Python web dashboard.

• Diabetes Classification Using Raman Spectrums

  Applied SVM, ANN, and signal processing techniques to classify diabetes from Raman spectrums, showing a strong foundation in data science fundamentals.

• Bros of Rainbow

  Conducted a comparative analysis of reinforcement learning architectures including Rainbow, DDQN, LSTM, and Transformer variants for OpenAI Gym game-playing agents.

• NLP Application

  Developed and hosted a web application showcasing multiple NLP models, including resume parser NER, sequence-to-sequence machine translation, and text generation.

• Serotonin Extraction and Quantitative Analysis

  Studied substance absorption with a monolith material approach for urine extraction to perform quantitative analysis of serotonin concentration related to depression.

Education