UPP Global Technology JSC

1. Overview

A big retail company in Vietnam faced challenges with their outdated data infrastructure, as it was scattered across multiple systems and tools/platforms. While it worked well when it was built, the growing demand in volume, variety, and velocity of data proved to be a challenge for said system. Our solution aimed to unify all data streams into one unified data platform for improved cost optimization, ease of governance, and future maintainability. 

2. Challenges

• Scaling Challenges: The system faced critical limitations due to a fragmented architecture split between batch and stream processing, lacking a dedicated real time database. Data silos across teams and between transformation layers and OLAP systems delayed insights. Scaling was costly and inefficient, especially with legacy systems like Apache Druid. Additionally, compliance with security regulations was difficult, and the overall infrastructure could not support modern workloads such as ML and AI.

3. Solutions

• Unifying the Data Platform
  We migrated all batch and streaming pipelines into a single Databricks Lakehouse. Databricks SQL replaced Druid for faster analytics, while Spark enabled scalable processing across data workflows.
• Powering ML & Real-Time Analytics
  ML pipelines and real time ingestion were built using Delta Live Tables and Spark. Governance was handled via Unity Catalog for secure, audit ready data access.

4. Outcome

• All in One Platform
  Migrated to Databricks for unified pipeline monitoring, analytics, and cost efficient scaling (all in one place).
• Faster & Simpler Analytics
  OLAP and real time analytics are now faster and easier with Databricks SQL and centralized data access.
• AI-Ready & Secure
  Enabled ML use cases and improved data governance with Unity Catalog and integrated model workflows.

5. Techstack

• Intergration
  •  Kafka
• Governance
  • MLflow
• Platform
  • Databricks
  • GCP
• Analytics
  • PySpark
  • MLFlows

1. Overview

A large hospital network implemented an on-premises AI assistant to aid clinicians in retrieving and summarizing medical knowledge from treatment guidelines, research articles, and patient records. This intelligent tool was designed to offer real-time, evidence-based recommendations during patient care, improving diagnostic accuracy, treatment consistency, and clinical decision-making across departments and specialties, while also reducing time spent on manual research tasks.

2. Challenges

• Inefficient Access to Critical Knowledge
  Clinicians faced increasing difficulty keeping up with the rapidly expanding body of medical literature, guidelines, and research findings, especially when making time-sensitive decisions during patient care.
• Manual Search and Privacy Risks
  Manual searches through medical databases were time-consuming and prone to error, while also posing risks to patient confidentiality due to inconsistent data access protocols and fragmented systems.

3. Solutions

• Private GPT-Powered Advisory Assistant
  A domain-specific large language model was deployed on-premises and fine-tuned with clinical guidelines and biomedical literature to ensure medical relevance and relevance and reliability in its responses.
• Secure EMR Integrated Retrieval Pipeline
  A custom retrieval system was integrated with the hospital s EMR and knowledge repositories, enabling the LLM to access patient records and reference materials in real time 

4. Outcome

• Proven Clinical Impact
  The pilot reported high accuracy in providing answers on par with medical experts, significantly reducing research time. Clinicians made faster decisions with ameasurable improvement in diagnostic efficiency, prompting plans for system expansion to additional departments and broader clinical use.

5. Techstack

• Model
  • Hugging Face
  • QLoRA, PEFT
  • bitsandbytes
• Retrieval
  • FAISS, Redis
  • Docling
  • LlamaParse
• Framework
  • Lang Chain
  • LlamaIndex
  • Gradirails
• Onpremise
  • K8s, Git, Minio
  • Jenkins, Docker
  • PLG, Airflow, Agro

1. Overview

A retail dropshipping chain faced challenges in managing its vast supply chain network. With millions of transactions, deliveries, and inventory updates occurring daily, identifying and responding to anomalies, such as delivery delays, stock discrepancies, and unusual sales patterns, was critical to maintaining smooth operations and avoiding losses.

The client required an AI-driven system to detect anomalies in real-time and recommend actionable insights to mitigate risks.

2. Challenges

• Supply Chain Complexity
  Fragmented vendors and warehouses made ops consistency hard to track.
• Data Overload
  Manual data checks were too slow to catch anomalies in real time.
• Customer Satisfaction Risks
  Delivery issues and stockouts hurt customer experience and brand trust.

3. Solutions

• Real-Time Anomaly Detection System
  Built real-time anomaly detection across transactions, deliveries, and inventory using unsupervised ML. (e.g., Isolation Forest).
• Product-Level Anomaly Insights
  Built item-level anomaly detection using sales, returns, and feedback to flag stock issues and quality risks.
• Root Cause Analysis (RCA)
  Used RCA to classify anomalies by root cause and flag high-risk cases for action.

4. Outcome

• Real-Time Visibility
  Improved supply chain visibility enabled faster anomaly detection and resolution.
• Reduced Stockouts:
  Minimized stockouts and delays to reduce losses and boost customer satisfaction.
• Actionable Insights:
  Provided insights into recurring anomalies to support long-term decisions. 

5. Techstack

• Python
  • Pycaret
  • Tensorflow
• AWS
  • Lamda
  • S3
  • SageMaker
• YOLO
• Databrick
• OpenCV
• Tableau
• Snowflake

1. Overview

Automate pre-consultation patient triage through intelligent questionnaire systems. Focus on user experience by delivering contextually relevant questions whiile avoiding redundancy. Leverage LLM technology to automate question generation workflows and provide final disease predictions. Integrate with EMR systems to personalize questions based on patient history, and continuously improve model accuracy through feedback loops and clinician validation. 

2. Challenges

• The Importance of Pre-Diagnosis
  The pre-diagnosis stage plays a crucial role in the healthcare journey. It sets the foundation for timely and accurate treatment by identifying potential health risks early on.
• Challenges in Traditional Processes 
  Previously, patients had to go through lengthy administrative steps and meet with general practitioners for initial assessments. Without clarity on their health status, many underestimated their conditions leading to delayed check-ups and missed early warnings. 

3. Solutions

• Smart Questionnaire System 
  Develop an application capable of providing diverse medical questionnaires and automatically selecting the most relevant questions to  identify potential causes of illness (Adaptive Question Algorithm). 
• LLM-Powered Question Generation 
  Utilize LLMs to automatically generate and compile question sets, allowing doctors to simply verify them instead of creating questions manually. 

4. Outcome

• Reduced pre-diagnosis time 
  The average time for patients to complete the pre-diagnosis process decreased by approximately 50%.
• Improved user satisfaction
  Over 80% of users rated the questionnaires as concise, easy to understand, simple to complete and expressed willingness to use the system again.  
• Seamless integration 
  Easily integrates with EMR/HIS systems via API.

5. Techstack

• Language
  • Python 
• Data storage 
  • Qdrant
  • Mlflow feature store 
• Deployment 
  • FastAPI, Docker 
  • vLLM
  • MLFlow 
• AI
  • scikit-learn 
  • shap, xgboost 
  • lightgbm 

Overview:

Client running businesses in the furniture field would like to convert their existing system from paper to cloud and develop a sales management system for their businesses in Viet Nam. In this project, we carried out the following tasks:

-Collecting information and expertise from all relevant departments.

-Gathering reports according to the requirements of relevant department heads.

-Collecting working habits and methods of relevant departments.

-Consulting solutions to convert from Excel to paper form by cloud.

-Implementing a Sales & Services system.

-Integrating available on-premium systems (Accounting, HR, etc).

Project's challenges and our solutions:
1. Every department has its own business, using its own excel format. Changing user habits from separate to common among departments becomes increasingly complex.

=> Divide implementation milestones, deploy each part to help employees adapt to each part, collect comments and suggestions to draw experience for the next deployment.

2. High operating costs

=> Use the solution to minimize the number of licenses needed

Outcome:

-The departments have access to the system in a short time.

-60% reduction in operating costs

Techstack:
Sales Cloud, Services Cloud, Salesforce Platforms ( Apex, Aura component, Salesforce process ... )

Overview:

The water flow warning system is developed with the following features:

1) To predict the water inflow of a wastewater treatment plant (Matlab/AI forecasting)

2) To set alarms and notification functions related to various system and sensor data, etc.

3) To collect, store, edit, and display on the web different data from sensor devices

4) To provide real-time sensor data acquisition and screen display

5) To notify the system admin if there is an abnormal detected in the sensor data.

Project's challenges:
Monitoring of water levels in the plant's reservoirs is done manually. This leads to slow response when situations occur. The customer has many treatment pools in different regions. It is desirable to make a highly accurate prediction of the water level of each lake.

Our solutions:
We propose to build a system running on Azure, allowing us to receive data from pre-installed sensors in lakes, and then combine it with the weather forecast in the corresponding location to predict the water level of the lakes. The data relating to the sensors is also stored so that training can be accomplished to improve the predictive model.

Outcome:
The system allows customers to predict possible situations in advance to make appropriate response plans. The prediction model is improved to give the best accuracy.

Techstack:

Azure, ReactJS, Spring Boot, Java, Highcharts

Overview:

To apply the corporate governance system of UPP Global Technology by building an enclosed system across Sales, Project Management, Human Resources, recruitment and Asset Management through which we can optimize license usage.

Project's challenges:

Salesforce is not ERP.

Our solutions:

For each small ERP system (Asset Management, Project Management, HR ....), Upp has brought the main objects of each of these systems such as asset, project, and human, etc to become customers in CRM. Take advantage of Salesforce's existing processes and flows as well as based on the specific needs of each system to create custom.

Outcome:

The system is built to meet the specific needs of each individual system.

Techstack:

Sales Cloud, Services Cloud, Salesforce Platforms ( Apex, Aura component, Salesforce process ... )

Overview:

As a credit company, they give thousands of approval/refuse to customers' applications. Clients want to analyze their decisions on which factor is the most important in the decision making. They also want to analyze the differences between years and prediction the change after each year.

Project's challenges:

Clients require a clear and detailed explanation of our output (generated by AI).

Our solutions:

We apply ExplanableAI so that the model could give some explanation for their reasoning. Furthermore, we analyze hundreds of the critical (represenatative and outliner) data points (customer profile). So that client is confident with our solution.

Techstack: PyTorch, MLOps (Kubeflow or Dataiku), Tableau

Overview:

After the consultation process, the next step is implementing and operating the data system. Here, we must ensure good system operation, high-quality data, and absolute security.

Project's challenges:

During the construction and operation we encountered the following problems:

1. We do not have in-depth knowledge of the technology used. Accompanied by big data problems, the processing performance (Ingest Data, ETL Data) has not been optimized in terms of processing time. Chaining the system again costs in terms of time, increasing a lot of data that is not updated in time, which will affect the following problems.

2. How to organize the design of pipelines, the accompanying components are overlapped, and the technology is not optimal. Increases the cost of system operating infrastructure.

3. Capturing business logic is not clear which leads to low-quality data output, incorrect information, more seriously not as required. Influence subsequent decisions or problems.

Solutions:

Encrypt sensitive data, and clearly decentralize the right people, the right authorities, responsibilities, and indirect access. Use technology proficiently, design clearly organize related component structures to avoid repetition. Optimized data query, and structure. There is a system to monitor and fix unexpected problems during data operation.

Techstack: Cloud, Programming Language, Database, Bigdata Frameworks, ETL Orchestration, Data Warehouse

Overview:

In this game, our responsibilities are as below:

- To participate in gameplay sketching and brainstorming process.
- To implement TBM token (ERC20) which is used for purchasing items in the game.
- To implement TBR token (ERC20) - a reward token that helps prevent game economy inflation.
- To develop game characters and items by using ERC1155 and ERC721.
- To build an upgradeable marketplace contract to help users trade game items.
- To implement a server API by using Moralis SDK.

Challenges:
-Inflation of the economy in the game
-High gas fee, slow game speed due to the influence of processing speed of blockchain network
-High gas fee when listing products as well as slow speed when querying and filtering products sold in the marketplace and transaction history

Solutions:
+ Issue 1 sub-token (TBR) for payout
+ Build mechanisms to burn both sub-token (TBR), main-token (TBM) and NFTs
+ Build a staking mechanism to prevent players from selling off user tokens
+ Save the metadata in the backend when only basic information is stored on the blockchain
+ Updated result information of each game is provided to the blockchain after the game is over
+ Save the metadata in the backend when only basic information is stored on the blockchain
+ Build lazy minting solution to transfer gas fee to user
+ Build event sourcing system to record and save transactions in the backend, which can help query and filter products in the marketplace and transaction history

Outcome:
-Reduced gas fee as well as increased speed and gaming experience that increases performance by more than 75%
-Reduced paid gas fee paid which boosts performance by query and filter by 80%
-The departments have access to the system in a short time.

Techstack: Solidity, Hardhat, Truffle, Moralis SDK, EtherJS, etc.