We used biological chemistry data measurements (proteins/peptides abundance) from Parkinson patients to understand and predict progression of the disease (using MDS-UPDRS scores), provide information about which molecules change as Parkinson's disease progresses.
Our work (public notebook kernel) got more than 45 upvotes and awarded a Notebook Silver Medal on Kaggle, subsidiary of Google LLC.
Parallel LSTM Training for Sequence Prediction from Sequential Data
We analyzed and paralleled the LSTM model (an RNN - Recurrent Neural Network) in order to improve its training speed and efficiency. We utilized NVIDIA GPU power for multi-threads parallel computing. The specific task that we applied for the demo is time-series prediction - sequence prediction from sequential data (seq2seq).
Our ParallelV3 is 1700% efficiency compared to Sequential version.
We introduce a bio‑inspired encoding framework for forecasting the direction of financial time series. Motivated by the limitations of linear models and the opacity of many deep learning approaches, we draw an analogy to genetics: observable micro‑patterns are encoded into symbolic "Financial DNA" sequences. These sequences are then analyzed using a probabilistic state‑transition mechanism to estimate the likelihood of subsequent market directions. We evaluate the approach on Bitcoin hourly OHLCV data with a rolling backtest. Among the horizons considered, modeling transitions from current Financial DNA patterns to the 4‑hour‑ahead price direction yields the strongest results, achieving a win ratio of 0.729. The findings suggest that compact, interpretable symbolic representations can capture salient, recurring structures in noisy, non‑stationary markets and support effective directional forecasts.
Design and develop the core of the information extraction system for land-use-rights-certificate digitization.
Design and develop the core of the AI assistant system for the Lam Dong Provincial Party Congress, term 2025-2030 (Đại hội Đại biểu Đảng bộ tỉnh Lâm Đồng nhiệm kỳ 2025-2030).
Design and develop the core of LLM information retrieval system for administrative procedure documents
National Public Service Portal (Cổng Dịch vụ công Quốc gia)
Ministry of Public Security (Bộ Công an)
Lam Dong Province (Tỉnh Lâm Đồng trước sáp nhập 2025)
Lang Son Province (Tỉnh Lạng Sơn trước sáp nhập 2025)
An Giang Province (Tỉnh An Giang trước sáp nhập 2025)
Design and develop the core of automatic scoring system based on computer vision for multiple choice tests
Department of Education and Training (Sở Giáo dục và Đào tạo)
Design and develop the core of machine translation system to translate between Vietnamese language and ethnic minority language
Tien Giang Province (Tỉnh Tiền Giang)
Ba Na (ISO 639-3: bdq) - Việt (ISO 639-3: vie)
Lam Dong Province (Tỉnh Lâm Đồng)
K'Ho (ISO 639-3: kpm) - Việt (ISO 639-3: vie)
Chu Ru (ISO 639-3: cje) - Việt (ISO 639-3: vie)
VNG Corporation
2022
Product Designer / Operator
Game-Entertainment (GE) / Global IP / MadPoly Studio
[1] Analyze using Gamer Motivation Model from Quantic Foundry, or Google Play's Player Types
[2] Different games, different genres, different research purposes... have different systems and mechanics to focus. For example, with PUBG MOBILE, focus on Weapons, Gears, Items, Vehicles system; while with Candy Crush Saga, focus on Game Difficulty, Pacing and Pinch Point.
[3] Includes: Economy/Currencies System, IAP Mechanics (focus on IAP Recommended Packs and Event Battle Pass), Rules and Vectors of Monetization, EV (Expected Value) Estimation, Target User Segment.
A/B Testing: Level Design: Re-design/Optimize current levels and design new levels that meet Master Roadmap requirements
LiveOps: Brainstorm and propose themes for PvP Seasons and Event Battle Pass. Analyze, brainstorm and propose solutions/changes/features to solve players' issues/needs from User Review Report by Customer Support team
Game Balancing: Adjust difficulty curve, balance source (tap) & sink, prevent in-game economy from inflation (but some inflation is fun ;))