Organisation: Trip.com Group (Ctrip), Shanghai, China

Description: An end-to-end evaluation system for LLM travel agents, owned independently — corpus construction, question generation, answer keys grounded in mechanically decidable attributes, automated answer-quality scoring, and leaderboards — built to produce rankings that survive re-runs, addressing the non-reproducibility of LLM-as-judge grading.

Highlights:

• Derived a query intent taxonomy (room type, restaurants, cancellation policy, …) from real user discussions via open coding with entropy-based saturation analysis, then applied engagement-weighted sampling to align benchmark coverage with the observed query distribution.
• Validated reliability through sample-size and ranking-stability analysis (130 questions × 5 models × 3 runs = 1,950 evaluations), establishing how many items are needed before the leaderboard ordering stops moving.
• Engineered a concurrent Python pipeline (async workers, Playwright, cascaded LLM filtering) that crawls, deduplicates, and validates real user discussions into a structured retrieval corpus.
• Preparing a first-author manuscript, with submission planned for late 2026.

Duration: Jun 2026 - Aug 2026

Tech Stack: Python | async Python | Playwright | LLM APIs | Evaluation & Metric Design | Statistical Experiment Design

Organisation: National University of Singapore, Singapore

Description: Co-designed a long-horizon evaluation benchmark for video world models, testing whether generated videos preserve evidence and maintain a consistent 3D world across camera departure, occlusion, viewpoint changes, and revisits. Manuscript in preparation for submission.

Highlights:

• Generated controlled synthetic data in Unreal Engine 5 (scripted camera trajectories, revisit paths, per-object segmentation) — chosen over in-the-wild video so that failures are attributable to the model rather than to uncontrolled scene variation.
• Contributed to designing quantitative metrics and evaluation protocols for reproducibly measuring these failure modes across models.

Duration: May 2026 - Present

Tech Stack: Unreal Engine 5 | Python | Video Understanding | Evaluation Protocol Design

Description: A profile-distilling, intent-aware shopping agent built for the TechJam 2026 Conversational E-Commerce Search Challenge. It turns a limited anonymous customer profile and a multi-turn dialogue into a ranked shortlist drawn from a frozen catalog of 50,000 Amazon products.

Highlights:

• Designed reversible conversation state that records constraint provenance, preserves factual answers, and retracts superseded preferences when a customer changes intent.
• Combined exact attribute intersection, fielded BM25, taxonomy-aware matching, and weighted reciprocal-rank fusion with intent-specific ranking and exposure-aware exploration.
• Developed a progressive recommendation policy that exposes one high-confidence result on the first two turns before opening the Top 10 as evidence accumulates.
• Improved the organizer BM25 starter from 0.125 to 1.000 Hit Rate@10; the final offline run found all 200 public targets, placed 190 at rank 1, and achieved 0.965853 MRR and 2.305 MTTC with zero model tokens.
• Built 79 automated tests, a deterministic 100-case unseen-target diagnostic suite, and a replay dashboard for inspecting conversation state, ranking evidence, and per-scenario metrics.

Project Links: Devpost | GitHub | Live Demo

Duration: Aug 2026 - Sep 2026

Tech Stack: Python 3.12 | BM25 | Reciprocal Rank Fusion | Structured Retrieval | Conversational State Management | SQLite | GPT-5.6 Luna & Qwen Embedding/Reranking

Description: A minimal RAG-based customer support agent backend that integrates retrieval, guarded prompting, and reproducible local deployment.

Highlights:

• Implemented RESTful endpoints with Swagger/OpenAPI for rapid testing and demo recording.
• Built a dev/prod configuration split using Spring Profiles (dev with H2; prod with MySQL + Redis).
• Added a Redis cache with TTL and graceful fallback, keeping the main request path available even when Redis is down.
• Containerized dependencies via docker-compose (MySQL + Redis with volumes and healthchecks) for one-command local setup.
• Maintained an architecture-first README (Mermaid diagrams and startup steps) aligned with the implementation.

Project Link: NetEase-music-agent-backend-demo

Duration: Dec 2025 - Jan 2026

Tech Stack: Java | Spring Boot | Spring Data JPA | Maven | H2 | MySQL | Redis | Docker Compose | Swagger/OpenAPI | OkHttp | OpenAI-compatible LLM API

Description: A hierarchical review filtering system using AI models we trained, routing requests across fast and accurate stages to separate genuine feedback from spam at scale.

Highlights:

• Designed a multi-stage pipeline: rule-based preprocessing → fast classifier routing → multimodal validation → RAG-backed final verification.
• Trained and evaluated text classification models (e.g., FastBERT/BERT-family) and integrated confidence-based routing to balance latency and accuracy.
• Incorporated multimodal signals (e.g., CLIP) to handle cases where text-only signals are insufficient.
• Supported both single-review and batch processing workflows for practical experimentation and benchmarking.

Project Link: Harmonia

Duration: Aug 2025 - Sep 2025

Tech Stack: Python | PyTorch | Transformers | FastBERT/BERT-family | CLIP | FAISS | Qwen | RAG | vLLM (inference) | OpenAI APIs

Description: An AI agent that plays Grid Adventure, a tile-based puzzle game, by planning optimal action sequences using A* search.

Highlights:

• Designed an A* search planner to compute optimal action sequences for navigating tile-based puzzle states.
• Supported two input modes: structured game state (Task 1) and raw pixel image perception (Task 2/3), requiring vision-based state extraction.
• Evaluated agent performance across varying grid complexities and input fidelity conditions.

Project Link: Grid Adventure Agent Github Repo

Duration: Jan 2026 - Apr 2026

Tech Stack: Python | Pytourch | A* Search | Image Processing

Description: A conceptual desktop system designed to act as a neutral, centralised Co-living Mediator for residents in high-density shared housing, completed as part of NM3243 Interaction Design at NUS.

Highlights:

• Conducted user research and developed personas, user journeys, and task flows to ground the design in real co-living pain points.
• Produced low-fi to high-fi prototypes in Figma, iterating based on usability feedback.
• Designed mediation and communication features to reduce friction between roommates in shared housing scenarios.

Project Link: HomeRun GitHub Repo

Duration: Jan 2026 - Apr 2026

Tech Stack: Figma | Miro | UX Design

Description: An individual portfolio for CS3240 Interaction Design at NUS, documenting personal contributions and design reflections on the MarketPlace service of PropBank — a super app built around a cosplay-rooted ecosystem.

Highlights:

• Contributed to the design and iteration of the MarketPlace service within a group UX project.
• Documented individual design decisions, rationale, and reflections aligned with Interaction Design principles.
• Applied user-centred design methods including wireframing, prototyping, and heuristic evaluation.

Project Link: PropBank Portfolio Github Repo

Duration: Jan 2026 - Apr 2026

Tech Stack: Figma | Next.js | TypeScript | UX & UI Design |

Description: A vehicle fleet management web app built for an automotive rental context, supporting real-time tracking and full CRUD operations on vehicle records.

Highlights:

• Built full CRUD operations on vehicle records including plate number, model, status, date, and notes.
• Implemented real-time search and status filtering (Available / Rented / Maintenance) with colour-coded indicators to support quick operational decisions.
• Applied React hooks (useState, useEffect) for state management and localStorage persistence, translating fleet management requirements into a clean, responsive UI.

Project Link: Fleet Dashboard Github Repo

Duration: Apr 2026 - Present

Tech Stack: React | Vite | JavaScript | localStorage

Description: A desktop task management application optimized for CLI users, combining the efficiency of command-line interaction with the clarity of a GUI. Built with strict adherence to Object-Oriented Programming principles for maintainable and extensible code architecture.

Highlights:

• OOP Design: Implemented comprehensive object-oriented architecture with proper encapsulation, inheritance (Todo/Deadline/Event extending abstract Task class), and polymorphism for command handling.
• Applied SOLID principles throughout the codebase, ensuring separation of concerns between UI, logic, and storage layers.
• Built three task types (Todo, Deadline, Event) with natural language datetime parsing and flexible scheduling capabilities.
• Designed command pattern implementation with aliases (e.g., todo/td, deadline/ddl) for rapid workflow execution.
• Implemented smart search functionality with keyword matching and snooze feature for rescheduling tasks.
• Integrated automatic data persistence with JSON serialization following proper abstraction layers.

Project Link: Mintty Github

Duration: Jan 2026 - Apr 2026

Tech Stack: Java 17 | JavaFX GUI | OOP Design Patterns

Description: A volunteer operations application adapted from an existing brownfield codebase to support the tracking and management of stray-cat records at NUS. The project focused on improving usability, aligning features with real user needs, and making the system easier to use and maintain.

Highlights:

• Adapted and extended an existing AB3 brownfield codebase into a context-specific operations app for volunteers managing stray-cat information.
• Conducted user research and developed user stories to better align the application’s features with actual operational workflows and user needs.
• Contributed to key functional enhancements such as add, find, edit, export, and other core CRUD-related operations.
• Improved the JavaFX user interface through layout refinements and CSS customisation, making the application clearer and more user-friendly.
• Updated major project documentation including the README, User Guide, and Developer Guide to support onboarding, usage, and maintainability.
• Worked within a structured team workflow using Git, branches, pull requests, merge reviews, weekly meetings, and release packaging for controlled collaboration and delivery.

Project Link: CatPals GitHub

Duration: Jan 2026 - Present

Tech Stack: Java | JavaFX | CSS | Git/GitHub | Brownfield Software Engineering

Description: A collection of OpenGL-based coursework demos exploring core real-time rendering concepts and interactive controls.

Highlights:

• Implemented graphics pipeline fundamentals, including transformations, camera controls, and scene organization.
• Built shader-based rendering effects (e.g., lighting, materials, and texturing) as part of course assignments.
• Added interactive input handling for navigation and parameter tuning during demos.

Project Link: OpenGL Coursework

Duration: Aug 2025 - Dec 2025

Tech Stack: C++ | OpenGL | GLUT

Description: A collection of coursework demonstrating 3D modeling theory, character animation, and computational geometry through Blender animation and C++/OpenGL mesh processing implementations.

Highlights:

• Created a collaborative short animation in Blender featuring custom character modeling, rigging with IK, keyframing with Bezier interpolation, and particle VFX systems.
• Implemented mesh processing algorithms including triangle mesh data structures (Trist), face/vertex normal computation, smooth shading, and boundary edge detection in C++/OpenGL.
• Built mesh navigation operations (enext, sym, org, dest, fnext) and component analysis using BFS/DFS for topology computation.
• Developed OBJ file parser with support for reading/writing Wavefront format, angle statistics computation, and interactive rendering modes (wireframe, flat/smooth shading).

Project Link: 3D Modelling and Animation

Duration: Jan 2026 - Apr 2026

Tech Stack: Blender | C++ | OpenGL | GLUT | Cycles Renderer