AI Systems, from Problem to Measurable Result.

01 / Workflow architecture

HR Travel Multi-Agent AI Workflow

An end-to-end multi-agent workflow project for conversational HR travel requests, policy validation, approvals, booking, cancellation, and modification.

• n8n
• Multi-agent orchestration
• RAG policy validation
• Microsoft Teams
• SharePoint
• Human-in-the-loop

My role

I handled the project end to end as a one-person delivery, translating HR travel requirements into agentic workflow logic, API integration design, RAG policy knowledge-base validation, human-in-the-loop approval controls, request-state management, and the supporting data and audit model.

How it works

Teams request -> intent routing -> structured extraction -> clarification and session memory -> RAG policy validation -> flight and hotel search -> preference filtering and ranking -> top-three recommendation options -> user selection -> manager-triggered real-time revalidation -> human-in-the-loop manager approval or rejection -> booking -> confirmation, status, cancellation, or modification.

Technical decisions

• Used a central orchestrator plus specialized worker workflows instead of one overloaded conversation flow.
• Used request IDs, execution IDs, persistent session memory, structured JSON handoffs, and explicit state transitions to support long-running operations.
• Designed API revalidation before approval and execution, idempotency controls, and Saga-style rollback for price changes, partial booking success, refunds, and modifications.
• Kept governance visible through authorization checks, role-based approval, audit records, policy versioning, and manual-review queues.

This project taught me that a useful agentic workflow needs more than orchestration: it needs clear states, policy checks, approval points, and recovery paths. I carried the design through the full request journey, with live supplier integration, credentials, testing, monitoring, and production controls as the next implementation steps.

02 / Cloud AI and agentic BI

Enterprise Conversational AI Platform

A paid-pilot enterprise platform for BFSI users to ask natural-language questions, generate governed analytics, surface insights, and schedule reports.

• Amazon Bedrock AgentCore
• AWS Lambda
• EventBridge
• RDS SQL Server
• QuickSight
• MCP tools

My role

Co-led architecture design, business-solution design, and stakeholder discussions in a 9-person team. I owned the end-to-end AI solution, agent and AWS cloud architecture, integration model, governance approach, and nine-sprint implementation roadmap, while the delivery team implemented the platform.

Architecture

Users submit questions through a web interface to Amazon Bedrock AgentCore Runtime. The Request Manager maintains conversation context, starts the workflow, and coordinates the implemented Data Analyst, Data Scientist, and Formatter agents. PostgreSQL stores application metadata separately from enterprise business data, while RDS SQL Server represents the connected enterprise data layer. Lambda, EventBridge, SES, S3, API Gateway, Cognito, IAM, EC2/ECS, CloudWatch, and QuickSight support the platform, with Power BI, Tableau, and Looker available as BI outputs.

What the system does

• Turns natural-language questions into governed SQL, descriptive analysis, forecasts, visualizations, executive summaries, and business-friendly responses.
• Lets specialist agents exchange intermediate results instead of forcing every request through one sequential workflow.
• Schedules recurring reports through EventBridge, Lambda, and SES, with MCP tools providing secure extension points for enterprise APIs and business systems.
• Keeps authentication, authorization, audit, logs, and runtime telemetry visible through Cognito, IAM, API Gateway, S3, CloudWatch, and Bedrock AgentCore.

Paid-pilot outcomes

• Supports 2 paid pilots across 5 business units and 20+ users, with capacity for up to 100 data sources and 30 recurring reports or workflows.
• Reduced typical ad-hoc analysis or report turnaround from 1-3 days, and sometimes more than 5 days, to initial fact-checked answers within 3 minutes.
• Saved 100+ analyst-hours per month across paid-pilot outcomes, with 95% adoption and 95% factual-answer accuracy after source validation.
• RM30k MRR remains a commercial target, not achieved revenue.

The paid-pilot work reinforced that conversational analytics only becomes useful when agent design, data boundaries, security, and business workflow are designed together. I learned to turn that architecture into a clear delivery roadmap while keeping data access, quality checks, and accountability visible. The public version keeps client details private.

03 / Data engineering and ML

Production ELT, forecasting, and BI delivery

A body of RapidMiner, Python, SQL, AWS, ClickHouse, Qlik Sense, and Power BI work covering production-oriented ELT, ML, and reporting.

• Python
• SQL
• RapidMiner
• ClickHouse
• XGBoost
• ARIMA

My role

Built and improved data preprocessing, ELT, modelling, evaluation, dashboarding, and automation workflows for operational and management reporting.

Approach

• Combined data preparation, quality checks, feature engineering, model evaluation, and BI delivery instead of treating them as isolated tasks.
• Used KNN, XGBoost, ARIMA, and other approaches according to the decision context and available data.
• Automated Excel, PDF, Word, ERP, email, and master-tracker workflows with UiPath to reduce processing time and improve reporting consistency.

This work took operational data through preparation, modelling, evaluation, dashboards, and automation so teams could use it for real reporting and decisions. I learned that dependable analytics relies as much on clean data, clear business questions, and repeatable delivery as it does on choosing the right model.

04 / Academic project

Airbnb new user prediction system

Final-year project at Tunku Abdul Rahman University of Management and Technology.

What I built

Python ETL, missing-value imputation, Z-score outlier detection, exploratory analysis, Streamlit presentation, and model comparison across XGBoost, Random Forest, Logistic Regression, and LightGBM.

Optimization

Compared GridSearchCV, RandomizedSearchCV, Bayesian Optimization, and Hyperband to improve model selection and ranking performance.

This project turned a 213,451-record dataset into a complete prediction workflow, from ETL and data cleaning to model comparison, optimization, and Streamlit presentation. I learned that model performance is only part of the job: the evaluation metric, optimization strategy, and final explanation all need to match the decision the system is meant to support.