I am Ewen Cheung Yi Wen, a Computer Science undergraduate at the National University of Singapore focused on software engineering, AI systems, and full-stack product engineering. My work sits at the intersection of backend systems, machine learning, LLM infrastructure, agentic workflows, data platforms, and production-grade user-facing applications.
I have built and deployed systems across RAG, text-to-SQL, vLLM inference optimisation, AI weather forecasting, private equity intelligence pipelines, multi-agent orchestration, academic planning systems, and pathfinding agents with CNN-based perception. I am also deepening my ML foundation across computer vision, diffusion models, generative modelling, and hallucination evaluation. I care about systems that are not only intelligent, but also reliable, auditable, secure, measurable, and usable by real people.
| Engineering Focus | What I Build |
|---|---|
| Software Engineering | Full-stack applications, backend APIs, database schemas, system design, testing workflows |
| AI / ML Engineering | RAG, text-to-SQL, model evaluation, inference optimisation, CNNs, GNNs, Transformers |
| Product Engineering | Responsive interfaces, user workflows, automation tools, end-to-end deployed products |
| Enterprise AI | Secure LLM-to-database workflows, Zero Trust-aligned access, monitoring, observability |
Open To
| Domain | Proficiency | Details |
|---|---|---|
| RAG Systems | Advanced | Hybrid retrieval, lexical/vector search, grounding, document/data pipelines, enterprise query answering |
| Text-to-SQL | Advanced | Real-time SQL generation, enterprise data grounding, database access controls, query evaluation |
| Agentic AI | Advanced | Tool-using agents, multi-agent orchestration, RBAC, approval-gated execution, workflow monitoring |
| LLM Infrastructure | Advanced | vLLM serving, batching, latency optimisation, Langfuse/LangSmith observability |
| Computer Vision | Strong | CNN-based tile recognition, 21-class classification, 210K+ augmented samples, constrained model deployment |
| Forecasting ML | Strong | GNN encoders, Transformer latent rollouts, observation-to-forecast modelling, meteorology workflows |
| Deep Learning Foundations | Strong | Transformer internals, attention mechanisms, encoder-decoder training, optimisation loops, PyTorch implementation |
| Generative AI Research | Developing | Diffusion models, hallucination evaluation, generative model reliability, multimodal AI systems |
| ML Evaluation | Strong | Evaluation pipelines, model quality measurement, investment analysis consistency, performance benchmarking |
| Full-Stack AI Products | Strong | React/Next.js frontends, FastAPI/REST backends, PostgreSQL/Supabase, responsive product workflows |
Jun 2026 - Dec 2026
Built an end-to-end private equity intelligence pipeline that transforms deal data into structured investment insights, accelerating underwriting analysis and improving evaluation consistency.
Scope of Work
- Developed evaluation workflows for private equity investment analysis.
- Automated repetitive underwriting and deal analysis processes.
- Generated structured, data-driven insights for investment decision-making.
- Connected software engineering, data processing, and evaluation consistency in a finance context.
May 2026 - Jun 2026
Developed an Aardvark-inspired end-to-end AI weather forecasting pipeline using GNN encoders and Transformer latent rollouts, converting raw meteorological observations into forecast-ready latent states for seconds-to-minutes inference.
Scope of Work
- Built observation-to-forecast modelling workflows for meteorology.
- Used GNN encoders and Transformer latent rollouts for forecast-ready representation learning.
- Reduced forecast initialisation latency from 3-6 hours to 5 minutes.
- Replaced traditional initialisation bottlenecks with end-to-end AI inference design.
May 2025 - Dec 2025
Engineered production-grade RAG and text-to-SQL systems for enterprise data analysis, with hybrid retrieval, real-time SQL generation, inference optimisation, and Zero Trust-aligned LLM-to-database access controls.
Scope of Work
- Achieved 90%+ accuracy on enterprise information queries.
- Built hybrid retrieval pipelines with lexical/vector search and enterprise grounding.
- Reduced LLM response latency by 30% through vLLM inference optimisation, batching, and serving workflow improvements.
- Improved auditability and access safety for enterprise data workflows.
One-Man Business Multi-Agent Orchestrator - Enterprise Agentic AI System
A supervisor-driven, role-aware multi-agent system for solo business operations. It combines a FastAPI backend, Next.js dashboard, Supabase/PostgreSQL data layer, Telegram/web messaging, LangGraph orchestration, retrieval, policy grounding, memory, external research, risk checks, and human approval gates.
| Category | Details |
|---|---|
| Stack | Python, FastAPI, Next.js, React, Supabase, PostgreSQL, LangGraph, Docker, Telegram Webhooks |
| Scale | 20+ relational tables, owner/stakeholder dashboards, multi-channel messaging, seeded business data, tests and evaluation workflows |
| Performance | Supervisor pipeline delegates work to retrieval, policy, research, memory, and reply agents instead of relying on one unconstrained LLM pass |
| Security | RBAC, role-aware tool/data access, policy grounding, risk evaluation, approval-gated execution, human-in-the-loop controls |
| Impact | Won Best Project Award with 24/25; demonstrates enterprise-grade agent harness engineering and operational AI system design |
| Repository | github.com/EwenCheung/One-Man-Business-Multi-Agent-Orchestrator-System |
This is my strongest project because it shows production-style software engineering around AI: orchestration, permissions, persistence, evaluation, UI, integrations, deployment, and safety boundaries.
AI-DOP - Aardvark-Inspired AI Weather Forecasting Pipeline
An AI weather forecasting research project inspired by Aardvark-style observation-to-forecast modelling. The repository includes encoder, processor, decoder, evaluation, end-to-end fine-tuning, WeatherBench-style evaluation, PBS training jobs, notebooks, and generated research outputs.
| Category | Details |
|---|---|
| Stack | Python, PyTorch, Transformers, NumPy, pandas, xarray, Matplotlib, Jupyter, PBS, MLflow |
| Scale | Encoder/processor/decoder training workflows, cached tensor pipelines, station evaluation, baseline preparation, and forecast evaluation scripts |
| Performance | Targets seconds-to-minutes AI inference by replacing slow forecast initialisation stages with learned observation-to-forecast components |
| Security | Local research pipeline with environment-based paths, reproducible scripts, and separated training/evaluation entrypoints |
| Impact | Bridges ML research and climate/weather infrastructure; directly aligns with NEA / CCRS meteorology work |
| Repository | github.com/EwenCheung/AI-DOP |
This is worth featuring because it is not just a concept repo: it contains real model-training, evaluation, notebook, and HPC-style workflow assets.
NUS Planner - Auto + AI-Powered Academic Planning Platform
An AI-powered academic planning platform that generates optimised 4-year study plans in seconds. The system combines automated module data pipelines, prerequisite enforcement, exchange mapping, mutually exclusive module handling, backend validation services, structured planning endpoints, and a responsive React frontend.
| Category | Details |
|---|---|
| Stack | TypeScript, React, Next.js, Supabase, PostgreSQL, pgvector, RAG, REST APIs |
| Scale | Multi-year academic planning across modules, prerequisites, semester availability, and exchange mappings |
| Performance | Generates optimised 4-year plans in seconds using a constraint-aware planning engine |
| Security | Environment-based configuration, private data support, backend validation |
| Impact | Turns manual academic planning into an automated, explainable, interactive workflow |
| Repository | github.com/EwenCheung/NUSPlanner |
This project represents my product engineering style: algorithmic planning, AI assistance, structured data, frontend UX, backend validation, and deployment readiness in one product.
CS2109S Path-Finding Agent with CNN and A* Search
A hybrid AI agent for Grid Universe tasks that combines classical A* search, CNN/RNN perception, and ML-based decoding. The system handles dynamic hazards, interactive elements, power-ups, multiple objectives, and constrained model deployment.
| Category | Details |
|---|---|
| Stack | Python, PyTorch, CNNs, RNNs, A* Search, Jupyter, custom evaluation scripts |
| Scale | 210K+ augmented image samples, 21-class tile recognition, multi-objective grid environments |
| Performance | 97/100 overall score, Top 5%, 100% tile-recognition accuracy under a 2 MB model constraint |
| Security | Offline local execution, reproducible environment, contained model artefacts |
| Impact | Demonstrates integration of search, planning, perception, ML modelling, and constrained AI deployment |
| Repository | github.com/EwenCheung/CS2109S-Path-Finding-Agent-with-CNN-and-A-Search |
This project is strong public proof of my ability to combine classical AI, deep learning, and practical engineering constraints into one working agent system.
Train a Transformer from Zero to Hero - English-Chinese NMT
A from-scratch Transformer implementation for English-Chinese neural machine translation. The project implements embeddings, positional encoding, scaled dot-product attention, multi-head attention, encoder/decoder blocks, training, checkpointing, BLEU evaluation, tokenizer preparation, and interactive translation.
| Category | Details |
|---|---|
| Stack | Python, PyTorch, custom Transformer modules, SentencePiece-style tokenization, BLEU evaluation |
| Scale | Web-crawled English-Chinese corpus, train/dev/test datasets, tokenizer pipeline, training and translation entrypoints |
| Performance | Tracks validation loss and BLEU score, saves best checkpoints, supports inference through an interactive translation script |
| Security | Local model training/inference pipeline with explicit configs and reproducible entrypoints |
| Impact | Demonstrates first-principles understanding of Transformer internals rather than only using high-level model APIs |
| Repository | github.com/EwenCheung/Train-a-transformer-from-zero-to-hero |
This project is worth mentioning because it shows model-level engineering: attention, encoder-decoder architecture, optimization, training loops, and inference mechanics.
SuperConfig - No-Code Agentic AI Configuration Platform
A no-code agentic AI configuration project built for the SimplifyNext x AWS hackathon. The platform focuses on rapid setup of personalised AI agents and workflow configuration for users who need agent capabilities without manually engineering the system from scratch.
| Category | Details |
|---|---|
| Stack | Python, AWS, agent configuration, workflow automation |
| Scale | Demo-oriented agent setup flow with configurable behaviour and deployment workflow |
| Performance | Designed for rapid configuration and fast agent setup |
| Security | Configuration-driven access patterns and controlled setup flow |
| Impact | Lowers the barrier to deploying personalised agentic AI workflows |
| Repository | github.com/EwenCheung/SuperConfig |
This project shows my interest in turning agentic AI from a research idea into a configurable product experience.
Dual Defence v3 - Python Tower Defense Game
A comprehensive Python/Pygame tower-defense game featuring campaign modes, progression systems, user management, custom game mechanics, and replayable gameplay loops.
| Category | Details |
|---|---|
| Stack | Python, Pygame, custom game logic |
| Scale | Multiple modes, progression mechanics, user management, custom gameplay systems |
| Performance | Local game loop with responsive interaction and state-driven mechanics |
| Security | Local user state handling and contained runtime |
| Impact | Demonstrates software design, game state management, event systems, and end-to-end project completion |
| Repository | github.com/EwenCheung/Dual-Defence-v3-latest |
This project demonstrates my foundation in software construction before moving deeper into AI and enterprise systems.
|
B.Sc. (Honours), Computer Science — Aug 2025 – May 2028
Data Science & Artificial Intelligence — Aug 2024 – Jul 2025
|
🥈 2x First Runner-Up — National Hackathons 🏅 2x Finalist — National Hackathons 🎖️ Google AI CTO Bootcamp — Selected Representative 🎯 Top 1-5% — NUS CS2109S (Score: 97/100) 🏆 Best Project Award — Multi-Agent Business Orchestrator (24/25) 🏫 NUSSU CommIT — Training Cell Head |
Learning:
- Advanced LLM evaluation and agent reliability
- Secure AI systems for enterprise data workflows
- Forecasting models with GNNs and Transformer latent rollouts
- Computer vision beyond classification: detection, segmentation, and representation learning
- Diffusion models and generative modelling fundamentals
Building:
- Private equity intelligence and evaluation pipelines at GIC
- AI weather forecasting pipelines with NEA / CCRS
- Full-stack AI products with reliable backend infrastructure
Exploring:
- Applied AI for finance and investment intelligence
- Agentic AI orchestration and tool-use reliability
- Hallucination detection, grounding, and evaluation
- Machine learning systems for CV, generative AI, and forecasting
- RAG quality measurement and text-to-SQL auditability
Open To:
- AI / ML engineering roles
- Software engineering internships
- Applied ML research collaborations
- Full-stack AI product projects
💡 "Coding is the process of building things from 0 to visible."