Rhys MirandaData & Systems Engineering, CityU Hong Kong

Rhys Miranda: Fraud, markets, and the odd qubit.

I'm an NVIDIA Student Ambassador in Hong Kong. Most of what I build sits where money meets messy data: fraud rings, market signals, and one detour into quantum error correction that I still think about.

See my work Get in touch

This is a simulated lending network. The pink clusters are fraud rings. Hover one to flag it.Tap one to flag it.

Open to 2026/27 graduate and internship roles in ML engineering, quant and data science Winner, HKSTP One Person Company (OPC) Hackathon 2026 Previously at Guns and Rain, Just Grace NPC and Suca Foods
Rhys Miranda speaking on stage with a microphone

Hi, I'm Rhys.

I study Data & Systems Engineering at CityU in Hong Kong. I grew up in Mumbai, and work and study have since taken me to South Africa, Vietnam and New Zealand.

When I'm not at a screen, I'm on a basketball court, near the sea, or halfway through a book. I try to leave every place, and every system, a little better than I found it.

Five countries, so far.

  • MumbaiWhere I grew up.
  • Hong KongHome now: CityU, NVIDIA, HKSTP Talent Foundry.
  • Johannesburg & LangaData and AI internships.
  • Buon Ma ThuotWebsites for Suca Foods with AIESEC.
  • AucklandExchange semester in cybersecurity and AI.

Things I've built.

Five projects, from GPU fraud pipelines to a quantum decoder. Each one started with a boring baseline it had to beat.

Money-mule network, simulated

NVIDIA, SME lending

GPU-accelerated fraud detection

Organised fraud in small-business lending hides in plain sight: each fake account looks normal on its own, so I look at the network instead. Real fraud data is private, so I built a generator for synthetic fraud graphs. At a realistic 1% fraud rate, a graph model found 2.7 times more fraud than a gradient-boosted baseline given the same information. At NVIDIA, that idea runs as a GPU pipeline pairing graph models with XGBoost, with agents reading streaming lending data so nobody has to hand-update rules.

2.7×more fraud than baseline
GNN + XGBoosthybrid model
Real-timeKafka and Morpheus

Built with RAPIDS, cuGraph, PyTorch Geometric, Triton, LangChain, NIM, SHAP

Overnight Indian trading hours
Silver ETF returns, 2022 to 2026

Market research, live pipeline

Does the news move silver ETFs?

I scored 234,000 headlines with FinBERT on Spark to see whether the morning's news predicts Indian silver ETFs. It doesn't: last night's silver and rupee moves already explain the day. The data showed something else instead. The ETF opened about 18 bps a day too high and gave it back within an hour, so buying at 10:15 rather than the open turned −50% into +15%, an edge that has since faded. A daily Airflow job now paper-trades it into MongoDB, behind a FastAPI service and a React dashboard.

234kheadlines scored
t = −4.0first-hour giveback
Livedaily paper trading

Built with Spark, FinBERT, BigQuery, Airflow, MongoDB, FastAPI, React, Docker, R

Live dashboardCode and write-up

Quantitative research

S&P 500 prediction and volatility trading

Markets are very good at making models look clever on data they've already seen. I engineered financial features, pitted ensembles against LSTMs, and used Monte Carlo simulation and risk-adjusted backtests to tell real signal apart from luck.

LSTMvs ensembles
Monte Carlosimulation

Built with Python, LSTMs, feature engineering, backtesting

Applied research

Quantum error correction with LLMs

Quantum computers make errors constantly, and a decoder has to guess which qubits went wrong. I had LLMs work alongside numerical optimisers to re-weight an MWPM decoder's matching graph, learning from the cases it got wrong. I report the two results separately, because simulation flatters.

2.8%lower error, simulated
0.74%lower error, Google hardware

Built with LLMs, BPOSD, MWPM, dual annealing, TuRBO

Photo Question CNN RAG LLM Plan

Full-stack GenAI

AI health and wellness assistant

Take a photo of your meal, ask a question, and get advice you can actually use. A CNN identifies the food, and a RAG pipeline turns that into dietary guidance in plain language.

95%food recognition accuracy
Deployedfull-stack app

Built with Flask, Hugging Face, RAG, OpenCV, SQLite

I don't trust a model until it beats a boring one.

Most of the job is checking my own work. A model can score well and still be learning the wrong thing. So I start with the dullest baseline I can find, and anything fancier has to beat it by enough to matter. If a graph neural net can't outdo gradient boosting on flat features, the graph isn't adding anything, and I'd rather find that out myself than hear it from someone else.

Baselines first

I report how much a model beats the simple version by, not just its headline score.

Show the working

If a system flags someone's loan, they should be able to see why and argue with it.

Numbers people feel

My proudest result in South Africa wasn't a model score. It was time saved for real people.

Then I try to explain the result to someone who wasn't in the room. If I can't, it isn't finished.

Experience.

Internships across three countries, then back to Hong Kong to build fraud systems with NVIDIA.

  1. NowFeb 2026 – present, Hong Kong

    Student Ambassador

    NVIDIA

    16 → 1source systems resolved into one risk record

    • Building a 3-agent LLM service (retrieval, fraud detection, compliance) for SME lending, with every agent's output checked against a schema before anything acts on it.
    • Built the team's risk data system: 16 source systems in 13 formats, 30+ risk indicators on 1M+ rows, moved to GPU for roughly an order-of-magnitude speed-up.
    • Built the evaluation framework the team tests model changes against, and attach SHAP explanations to every scored record.

    RAPIDSPyTorch GeometricXGBoostLangChainNIMSHAP

  2. Jun – Aug 2025, South Africa (remote)

    Data & AI Specialist

    Guns and Rain

    40%faster data retrieval for 15+ people

    • Migrated and standardised 2 database systems with validation rules and data dictionaries.
    • Deployed an LLM tool that turns newsletters and social posts into structured market-trend records, replacing a manual weekly review.

    PythonSQLLLMsData quality

  3. Jun – Aug 2024, South Africa

    Data Scientist

    Just Grace NPC

    20+staff and managers trained in data and Power BI

    • Integrated 4 cloud data sources into one Power BI reporting layer.
    • Built dashboards that let senior management track Key Risk Indicators in real time.

    Power BICloud dataTraining

  4. Jun – Aug 2024, Buon Ma Thuot, Vietnam

    Website Developer

    Suca Foods, with AIESEC

    • Led development of the company's websites, improving how its products and services are presented to customers.

    HTMLCSSJavaScript

What I work with.

Python first, C++ when speed matters, and Spark or GPUs when the data gets big. Here it is as a stack, from the code up to what it builds.

57tools and methods
5layers of the stack
  • Used in the projects on this page
  • Also comfortable with
  1. What I buildLLM apps, quant and research
    LLMs and multi-agent systemsRAGLangChainNVIDIA NIMHugging FacePrompt engineeringMonte CarloBacktestingWalk-forward evaluationVolatility modellingTransaction cost analysisQuantum error correctionMWPM and BP-OSD
  2. ModelsHow it learns
    PyTorchPyTorch GeometricXGBoostSHAPFinBERTscikit-learnTensorFlow / KerasCNNsLSTMsOpenCVFeature engineering
  3. Compute and servingWhere it runs
    RAPIDS cuDF, cuML, cuGraphTriton Inference ServerMorpheusFlaskFastAPIgRPCDockerVercelGitHub ActionsAWSNeMo GuardrailsGit
  4. DataWhere it lives and moves
    Apache KafkaSQLiteApache SparkBigQueryApache FlinkApache AirflowHadoopPig LatinHiveQLMongoDB and NoSQLPower BITableau
  5. LanguagesWhat I write it in
    PythonSQLC++RScalaJavaScriptNode.jsReactHTML and CSS

Life outside the terminal.

The habits that keep the rest of it going.

An open book held up against misty forested hills

Always mid-book

There's usually a paper open in one tab and a novel on Adobe. Non-fiction on weekdays, fiction on weekends.

Basketball players celebrating on a court

Pick-up games

Nothing clears a stubborn bug like a couple of hours on court. Happier making the assist than taking the shot.

A curving beach and coastline seen from a hilltop

Coasts & Clear Water

The ocean was never far away in Mumbai or Hong Kong. New Zealand's coastline set a high bar.

Collage of records, headphones and musical instruments

Guitar and a playlist

I play guitar and sing, mostly just for myself. What I listen to depends on my mood: something to wind down, something to lift it, often on the commute.

Table tennis paddles and balls on a wooden board

Table tennis

I play purely for fun, and it's one of the few things that makes me lose track of time.

Clouds at sunset seen from an aeroplane window

Travelling

My favourite part of every new place is the people: the meals shared, the culture that shapes them, and the new perspectives.

Hiring? Or just want to talk?

I'm looking for 2026/27 graduate and internship roles in ML engineering, quant and data science. I'm also happy to talk fraud graphs, order books, or basketball. Email is the quickest way to reach me.

Email me LinkedIn GitHub Download CV