Building intelligent
software from idea to production.
I'm a fresh CS graduate specialising in Machine Learning and AI. I like building things end-to-end, from quick prototypes to polished, deployed products that real people use.

A full-stack engineer who ships real software.
I'm a recent Computer Science graduate from SIM UOL, where I earned First Class Honours specialising in Machine Learning and AI. I like working across the whole stack, from the database up to the interface, and seeing how the parts fit together.
I got my start at NEC Asia Pacific, where I helped migrate a legacy enterprise system to Azure. Later, I came back on contract to rework the SQL schemas and fix bulk-insert bottlenecks for over 100,000 records. Tracking down why those large uploads were so slow, and fixing it at the query and schema level, taught me that how you structure the data matters as much as the code that uses it. I also pushed hard to keep my academic performance consistent, earning the IRAS Book Prize (Gold in 2024, Silver in 2025).
Outside of work, I build end-to-end applications. My recent projects range from Trackery (a custom React/Node.js desktop inventory system) and TogetherSafe (a cross-platform emergency app) to a fine-tuned BERT classifier and a RAG app I built from scratch. I like working across the entire stack because it lets me take an idea and ship it as a complete, polished product.
- 2025IRAS Book Prize — Silver AwardIRAS
- 2024IRAS Book Prize — Gold AwardIRAS
- 2020vSplash Awards — 1st Runner-upSingapore Computer Society (SCS)
- 2020PolyFinTech100 API Hackathon — 1st Runner-upPolyFinTech
- 2018Director's Roll of HonourRepublic Polytechnic
Projects I'm most proud of.
Each of these tackled a different problem, from inventory systems to ML classifiers, and taught me something I couldn't have learned from coursework alone.
Trackery
A desktop inventory system I built for a family business.
A desktop inventory app I built for a family business, replacing their pen-and-paper purchase orders and sales tracking. I designed the PostgreSQL schema around how they actually work day-to-day, then built the React and Node.js app to be something they could rely on without a developer on call.
- ▸Built the frontend with React and Electron, wired to a remote PostgreSQL database so inventory stays in sync in real time.
- ▸Set up a zero-trust network with Tailscale so the business could access the system from anywhere without exposing the database to the public internet.
- ▸Replaced a manual pen-and-paper workflow with a proper system they can trust for daily purchase orders and sales tracking.
Let It Cook
An AI research agent that plans its approach, picks the right tool for each step, and checks its own work, with every step shown.
My follow-up to Pull The Receipts, moving from retrieval into agents. Instead of answering in one pass, it plans an approach, picks between web search, document retrieval, and a calculator, then reviews its own draft before answering. I built the reasoning loop by hand in LangGraph so every step is visible, and the interface streams it live as the agent plans, picks tools, and corrects itself.
- ▸Built the agent in LangGraph around the four core agentic patterns: planning, tool use, memory, and reflection, all as a custom stateful graph.
- ▸Gave it three tools and let it choose between them with function calling: web search, document retrieval over local embeddings, and a calculator parsed from an AST so it can't run arbitrary code.
- ▸Added a self-correction loop where the agent grades its own draft and revises when it falls short, with hard caps so it always terminates.
- ▸Streamed the agent's reasoning live to a Next.js frontend, with rate limiting to keep API costs in check and LangSmith tracing for observability.

Pull The Receipts
Source-grounded document Q&A that won't answer without showing its receipts.
A retrieval-augmented generation app that answers only from the documents you give it, citing the exact source behind every claim, and saying so when the answer isn't there rather than guessing. I built the whole RAG pipeline by hand instead of reaching for LangChain, because I wanted to understand and explain every stage rather than hide it behind a framework.
- ▸Wrote the full RAG pipeline from scratch: document loading, sentence-aware chunking, local embeddings, vector retrieval, prompt augmentation, and generation, with every step explicit.
- ▸Grounded every answer in retrieved chunks with inline citations to file and page, and made the model refuse when the documents don't back up an answer.
- ▸Put the embedding model, vector store, and LLM provider behind clean interfaces, so each is swappable and the test suite can fake all three to run offline.
- ▸Streamed answers token by token over Server-Sent Events with sources sent first, then containerised the backend with Docker for deployment.

Fake News Classifier
Fine-tuned BERT with explainability, deployed end-to-end.
A full-stack fake-news classifier built on a fine-tuned BERT model. Paste a statement and it returns a verdict, confidence, and the exact tokens that drove the decision. I compared five model architectures along the way, from Naive Bayes to BERT with metadata, to find where the real accuracy ceiling sits.
- ▸Fine-tuned bert-base-uncased on the PolitiFact dataset and compared five architectures, finding the accuracy ceiling sits with the task itself, not the model size.
- ▸Added token-level explanations with integrated gradients (a more reliable signal than raw attention), so you can see which words pushed the prediction.
- ▸Split it into a real microservice setup: FastAPI backend on Hugging Face Spaces, Next.js frontend on Vercel, model on HF Hub, each deployable on its own.
- ▸Surfaced a finding most tutorials miss: the bottleneck was missing context, not capacity. Adding speaker and source metadata lifted accuracy from 69% to 71%.

Howlennz Portfolio
A fast, self-managed portfolio for a Live2D artist.
Lennz is a Live2D artist who was stuck on a slow Carrd page with no way to manage his own content. I rebuilt it as a Next.js site backed by Sanity, so he can update his gallery, commissions, and about page himself without touching code.
- ▸Rebuilt a slow Carrd page as a Next.js 14 site that loads far faster, giving Lennz a more professional first impression for commission clients.
- ▸Designed the Sanity content model so he can manage his gallery, work history, and contact details himself, with no developer needed for updates.
- ▸Hit a Cloudflare Pages and Sanity edge-runtime incompatibility mid-deployment, debugged it, and migrated to Netlify without compromising the stack.

TogetherSafe
Cross-platform mobile app helping communities coordinate during emergencies.
My final year project, built solo: a mobile app to help communities coordinate during local emergencies, with shared alerts and check-ins. I built every layer myself, from the UI to the backend, and treated it like a real product rather than a school assignment.
- ▸Wired up REST APIs to push real-time emergency alerts and safety updates straight to users.
- ▸Built the interface in Flutter and Dart, focusing on clarity so people can find critical information instantly during stressful moments.
- ▸Owned the whole stack myself: UI, frontend, and backend, end to end.

The journey so far.
BSc (Hons) Computer Science, First Class Honours — Machine Learning & AI
Graduated with First Class Honours, focusing my studies on Machine Learning and AI while building full-stack and mobile projects on the side, like my Flutter capstone, TogetherSafe. I was also awarded the IRAS Book Prize twice (Gold in 2024, Silver in 2025) for academic performance.
Assistant Engineer
NEC brought me back on contract to fix the database bottlenecks we hit after the initial launch. I rewrote the SQL queries and adjusted the schema to handle massive data uploads smoothly. I also built a companion Flutter app so the staff could check inventory straight from their phones.
Application Developer — Intern
I led a small intern team to build a C# inventory system from scratch, replacing the department's old Excel sheets in just five months. We handled the entire process ourselves, from gathering requirements from the admin staff to deploying the system on Azure and loading in 100,000+ rows of legacy data.
Diploma in Information Technology
This is where I got my start in programming. Beyond my classes (where I made the 2018 Director's Roll of Honour), I spent a lot of time testing my skills in hackathons. I took 1st Runner-up at both the PolyFinTech100 Hackathon and the vSplash Awards, where I pitched an early computer vision prototype.
Let's build something
great together.
Whether it's an ML idea, a full-stack build, or just a weird side project you want a second pair of eyes on, I'd love to hear about it.
riskilinardi@gmail.comUsually replies within a few days · riskilinardi@gmail.com