- Systematically evaluate the performance of Small Language Models (SLMs) with less than 2B parameters on mental health text analysis. Fine-tuned SLMs to achieve on-par performance with zero-shot ChatGPT and studied the effect of quantization on the tuned models. The report is publicly available. - Measured and analyzed the efficiency of SLMs including memory footprint, processing and generating throughputs on Raspberry Pi 5 with GGUF. Made a contribution to llama.cpp open-source project to support the used mode

더보기

- Studied the performance of different Vision Transformer embedding patch sizes using the 2022 AICC traffic camera dataset. - Engineered the fine-tuning and evaluation pipeline for car detection only and sparse evaluation. Optimized the evaluation processing times 12x faster with multiprocessing and pre-filtering.

더보기

- Replaced a manual process with an automated pipeline using Amazon Textract to accurately and scalably extract information from PDF documents. Applied linear regression to rescale the extracted data and match it to the corresponding Acrobat fields in the original template. This reduced the time required to process a hundred-page document from one week to approximately two days, achieving a 70% reduction in processing time. - Built a standalone demo application of extracting pipeline, allowed employees to process documents at large scale and customizable filter for better matching performance. Set a foundation for a new service offered by the company. - Built a document management tool with storage, permission, and audit logs features. This tool allows users collaborate with company employees to manage documents, reuse document information within the tool and other products. - Collaborating with design team and product team to optimize and improve user experience of current products. - Technology: Frontend (Scala.js Laminar & React), Backend (Scala ZIO, Protobuf, FoundationDB, GraphQL, Temporal Workflow).

더보기

- Involved in the Looina project: a dynamic, highly customizable document tool that enables users to create various types of components within a document and view the document in multiple formats for optimal usability and support - Optimize the render process with lazy loading mechanism and ensure the lazy loading mechanism compatible with current features such as draggable sort and searching feature. Eventually, new lazy mechanism smoothened scrolling process of hundred-component document without latency - Technology: Frontend (Vanila Javascript, Re:dom)

더보기

- Collaborated with fellows, from Engineer to Business, and managed to achieve 500.000 downloads in total on App Store and Google Play Store. At peak, Luveri was on 5-th place in Social Networking in App Store Vietnam. In addition, #luveri hashtag got millions views on Tiktok and went viral in Vietnam, Thailand, and Philippines. - In charge of key features implementation (Home, Chat, Timeline/Calendar, Profile, Notification, Theme), including designing data schema, writing backend API, and building frontend. - Highlight implementation - Draggable widgets on home page, allowing users to customize their home interface by choosing the design they like and putting in the location they want - High usability chat interface with reaction, gif, stickers, voice messages, and detailed information - Calendar single day interface, connected with other features to allow users review their activities (chat logs, uploaded photos, written notes) on specific day. - Technology: Frontend (Flutter/Dart), Backend (Firebase, Nodejs, Isar, Hive).

더보기