Ted Chaiwachirasak

@Sidekick - Fine-tune large language models to improve conversational accuracy and relevance - Improve the evaluation to measure model performance - Design data pipelines to facilitate efficient fine-tuning experimentation

- Developed an LLM-powered conversational search assistant. Utilized Retrieval-Augmented-Generation (RAG) to provide similarity-based demonstration selection. - Leveraged in-context learning with few-shot prompting to implement guard rails tailored to the search use case. - Developed and optimized recommendation ML models for various e-commerce clients, overseeing model research, selection, and deployment to ensure positive ROI and meet latency requirements. - Implemented a custom Python framework from scratch to streamline analytics report sharing with customers and internal teams, enhancing communication and collaboration across the organization. - Developed operational toolings to provide real-time visibility into system health and recommendation metrics (Google Cloud Platform Dashboard, Streamlit).

- Reduced the compute time for the existing Hyperparameter Optimization for recommendation models by 59% through meta-learning (RGPE) and multi-fidelity optimization (BOHB).

- Constructed Hotel2Vec embeddings from user behavior data (TensorFlow) to improve the existing hotel recommendation system, achieving a statistically significant uplift in bookings as measured by an A/B test. - Applied the LISTwise ExplaiNer (LISTEN) algorithm to extract actionable insights such as feature selection from hotel ranking results generated by the deep recurrent neural network model (Bi-LSTM). - Built a React.js web app to allow the team to evaluate hotel ranking results interactively, streamlining the development and evaluation process.

- Enriched user experience by leveraging user-generated data to create customized content such as hotel recommendations, personalized hero images, and pre-selected hotel filters. - Deployed large-scale ML systems that serve millions of daily users, ensuring high performance and reliability while maintaining a seamless and engaging user experience. - Identified bottlenecks in the hotel recommendation service by analyzing terabytes of traffic data (Apache Hive, Impala). - Optimized the hotel recommendation model’s data size using Scala's standard serializer instead of MLeap, drastically reducing the service’s uptime by 80%. - Set up the CI/CD process to automate docker deployment, integration test, and load test on TeamCity, shortening the deployment process.

- Developed an end-to-end character embedding convolutional neural networks model for capturing character-level misspelling in Thai social media text. - Used Word2vec model for word embedding and clustering to capture spelling variations of words with the same meaning. - Implemented a baseline model for Thai real-word error detection using local word N-gram. - Augmented Thai dataset for non-word error detection using Gaussian distribution of edit distance from keyboard to mimic mistyping words.

- Implemented a recurrent neural network model with TensorFlow for sentiment analysis on social media text data obtained through Facebook Graph API. Obtained f1-score of 74% on 13k rows of labeled data. - Created a real-time labelling tool for DTAC’s call center team to collect labelled sentiment data. - Implemented an interactive dashboard for the business team to explore social media trends using Python Dash and Plotly. - Researched and compared Thai text segmentation tools in Python on InterBEST 2009/2010.

- Certificate of Academic Excellence for achieving the Second Highest Academic Rank among second-year students of Computer Engineering curriculum. - 3 Year continuing Scholarships for Academic Outstanding Students.

- Received scholarship from Erasmus. (European Union for exchange student)