Manchester, UK · Data Science & AI

NicoleMascarenhas

I'm a Data Science Master's student at the University of Manchester. I love exploring data, finding patterns in it, and turning those insights into working ML systems: predictive models, RAG pipelines, and more.

Global Futures Scholar Graduating Sep 2026 Open to opportunities
About

Hi, I'm Nicole!

Data is everywhere, but raw data on its own means nothing.

What interests me is the process of making sense of it: exploring it, finding the patterns hidden inside, and building something useful from what you find.

My dissertation fuses patient demographics, timestamped vitals, clinical events, and clinical notes from the MIMIC-III ICU database to jointly predict mortality, length of stay, and clinical phenotypes. ICU data is noisy, incomplete, and multimodal, and that complexity is precisely what the architecture is designed to handle.

Before Manchester I completed a First Class BEng in AI & Data Science in Mumbai, which gave me the foundation to build on.

MSc Data Science
Business & Management
University of Manchester, graduating September 2026
BEng AI & Data Science
First Class Honours
University of Mumbai, with Honours in Cybersecurity
🏅
Global Futures Scholarship

Merit‑based scholarship awarded by the University of Manchester for academic excellence.

Capabilities

Skills

What I actually use when building things!

ML & Deep Learning
PyTorch · TensorFlow / Keras · scikit-learn · XGBoost LSTM / GRU · Transformers · Cross-Attention · GANs · CNN / VAE · H2O GLM · Isolation Forest · SHAP
NLP & Gen AI
RAG Pipelines · LangChain · Prompt Engineering FAISS · Gemini API · Qwen2.5 · BGE / Sentence-Transformers · BM25 · Cross-Encoder Reranking · spaCy · TF-IDF / GloVe
Data Engineering
Python · SQL · R · MongoDB · Pandas / NumPy Docker · ETL Pipelines · Django REST · Flask · Selenium · BeautifulSoup · Snowflake · BigQuery · PySpark / MLlib
Analysis & Visualisation
Power BI · Statistical Analysis · Excel Plotly · Matplotlib / Seaborn · Tableau · D3.js · ARIMA / Time Series · A/B Testing · Hypothesis Testing
Tools & Deployment
Streamlit · Git / GitHub · Power Automate Gradio · Azure ML · Jupyter · PyInstaller
Work

Projects

Things I built because a problem was interesting, or a dataset was too good to ignore, or I just wanted to understand how something worked.

All projects
MIMIC-III · EHR · Multimodal Fusion
Dissertation 2026 — ongoing
Multimodal Learning for Disease Risk Prediction
MIMIC-III is a de-identified ICU database from Beth Israel Deaconess Medical Center in Boston, 58,000 admissions, over a decade of critical care records.

My dissertation builds a model that reads all four data types together: patient demographics, timestamped vitals, clinical events, and free-text doctor notes. LSTM and Transformer encoders handle each stream, connected by cross-attention, then jointly predict mortality, length of stay, and clinical phenotypes in one pass.
In progressArchitecture designed to jointly predict 3 clinical outcomes — mortality, length of stay, and phenotypes — not yet finalised or evaluated.
PyTorchMIMIC-IIITransformersCross-Attention
! 0.86 PR-AUC SDK.finance · Industry-Mentored
Industry-Mentored Feb – May 2026
Fraud Detection for SDK.finance
Developed in collaboration with SDK.finance as an industry partner, on 284K real credit card transactions and 6.3M PaySim synthetic records.

Two-stage pipeline: Isolation Forest flags the suspicious top 10% first, then H2O GLM scores each one with an interpretable probability. Financially interpretable features were engineered to avoid data leakage. Deployed end-to-end as a Streamlit investigation console with tiered risk decisions (approve / monitor / block).
77%fraud recall at PR-AUC 0.86, cutting the review queue by 90%
Isolation ForestH2O GLMscikit-learn
BKT IRT Spaced Rep. A/B tested · Flask · MongoDB · D3.js
Full Stack Jul 2024 – May 2025
AI-Enhanced Learning Platform — LearnTube
Turns any YouTube video into a personalised study session — quizzes, flashcards, summaries, and a concept map — driven by three adaptive models running in parallel per learner.

Bayesian Knowledge Tracing treats mastery as a hidden variable and updates it after every answer. A 3-parameter IRT engine selects each next question by maximising Fisher information at the learner's estimated ability. Three competing forgetting-curve models — Ebbinghaus, ACT-R, and a personalised BayesianRidge — race to predict the optimal review date, scored on held-out MAE/RMSE. Teaching-strategy experiments run through a proper A/B pipeline: power analysis before launch, Holm–Bonferroni correction, and Cohen's d effect sizes.
29unit tests covering the adaptive engine, plus a Chrome extension for in-browser quizzing
FlaskMongoDBGemini APIIRT / BKTD3.js
SHAP waterfall · per-patient explanations Random Forest · explainable AI
Explainable AI Jun – Aug 2023
Disease Outcome Predictor (SHAP)
Built during my internship at Tannmann Foundation. It predicts patient disease outcomes from symptom profiles (fever, breathing difficulty, blood pressure, cholesterol).

Every prediction is explained using SHAP TreeExplainer with per-patient waterfall charts. Deployed as an interactive web app with global feature importance and a per-patient data explorer.
100%of predictions explained via per-patient SHAP waterfall charts
PythonRandom ForestSHAPscikit-learn
450+ docs Wikipedia BGE FAISS + BM25 0.6/0.4 Cross-Encoder Reranker Qwen2.5 CoT Prompt RAG · Hybrid Retrieval · South Asian Cuisine QA
Gen AI Feb – May 2026
RAG Cuisine Agent
A full retrieval-augmented generation pipeline for South Asian cuisine QA. Scraped 450+ docs from Wikipedia, Wikibooks, and food blogs with adaptive chunking per source type.

Hybrid retrieval: BGE-large dense vectors (FAISS) weighted 60% + BM25 sparse 40%, then re-ranked by a cross-encoder before hitting Qwen2.5 with chain-of-thought prompting.
450+docs indexed with hybrid dense + sparse retrieval, cross-encoder re-ranked
FAISSBGEBM25Cross-EncoderQwen2.5
EY Data Challenge 2026 Snowflake · scikit-learn End-to-end ML pipeline Open Science Data Challenge · Team competition
Hackathon 2026
EY Open Science Data Challenge 2026
Given ~200 river locations across South Africa, we had to predict water quality: Total Alkalinity, Electrical Conductance, Dissolved Reactive Phosphorus, for completely unseen locations.

Features came from Landsat spectral bands, TerraClimate climate variables, land cover fractions, and terrain data. We trained CatBoost, LightGBM, and XGBoost with Optuna tuning and built an ensemble.
Key detailDiagnosed spatial leakage via SHAP before finalising the CatBoost/LightGBM/XGBoost ensemble for unseen locations.
CatBoostLightGBMXGBoostOptunaSHAPLandsat
Bi-LSTM 2M+ seqs GloVe 300d Attention custom Softmax next word temp sampling Streamlit live demo · Classic literature corpus
NLP Jan – May 2024
Next Word Predictor with LSTM & Attention
Language model trained on 2M+ sequences from classic literature: iterated from vanilla LSTM, to bidirectional LSTM, to a custom attention mechanism, to GloVe 300-dim transfer learning. Temperature sampling lets you dial the creativity up or down. There's a live Streamlit demo you can actually try.
2M+training sequences, with GloVe transfer learning and a custom attention mechanism
TensorFlowBi-LSTMGloVeAttention
optimal risk → return → –8.85% → –3.10% with bond allocation
Quant Finance 2025
Portfolio Risk & Optimisation Engine
Downloads live market data and calculates risk metrics: VaR, CVaR, Sharpe ratio, and maximum drawdown. Runs Monte Carlo simulation to construct the Efficient Frontier and find optimal allocations. Benchmarks against the S&P 500 (alpha and beta), and stress-tests against crash scenarios.

Key finding: adding a 40% bond allocation cut simulated Tech Meltdown losses from –8.85% to –3.10%.
65%smaller simulated crash loss (–8.85% → –3.10%) with a 40% bond allocation
PythonNumPyPandasyfinanceSciPy
639 locations · 31 states · SDG 6 WHO / CPCB threshold benchmarking
Data Engineering Jun – Aug 2024
India Water Quality Dashboard (SDG 6)
Built during my CSRBOX internship.

An interactive Streamlit dashboard analysing water quality data across 639 monitoring locations in 31 Indian states. Automated A–F grading against WHO and CPCB thresholds, Plotly visualisations, and EDA on BOD, dissolved oxygen, pH, and fecal coliform to identify pollution patterns.
639monitoring locations across 31 Indian states, auto-graded against WHO / CPCB thresholds
PythonStreamlitPlotlyPandas
engagement prediction · PySpark EDA 100K posts · XGBoost R² 0.9992
ML · Big Data 2025
PySpark Social Media Engagement Predictor
Distributed EDA and feature engineering in PySpark on 100K social media posts, feeding a production scikit-learn pipeline.

Benchmarked a Linear Regression baseline against XGBoost and LightGBM, deployed as a real-time Streamlit prediction app.
R² 0.9992achieved by XGBoost vs. 0.693 for the linear baseline
PySparkXGBoostStreamlit
▶ "open browser" ✓ done ▶ "take screenshot" Packaged .exe · Gemini NLU · Async tasks
Gen AI Jul 2023 – Apr 2024
Cypher — Desktop Voice Assistant
Desktop voice assistant packaged as a standalone .exe, no Python install needed. Voice commands trigger document creation, web search, screenshots, and system tasks. Gemini API handles the natural language understanding.
Key detailOne voice interface covers four different task types — documents, search, screenshots, and system commands — instead of a single-purpose skill.
Gemini APISpeechRecognitionTkinterPyInstaller
person conf: 0.91 → Firebase SSD MobileNet · Raspberry Pi · MAVLink · Firebase
Edge AI 2023 – 2024
Autonomous Drone Surveillance System
Built as part of my college drone club's entry into the SAEINDIA Aerothon. I was responsible for the object detection and tracking system: the drone's camera feed goes into SSD MobileNet for inference, then a centroid tracker assigns stable IDs to detected objects across frames.

The tracker compares each centroid to the frame centre to determine where the target is relative to the drone.
91%detection confidence, tracked with stable IDs across frames via centroid tracking
TFLiteOpenCVRaspberry Pi
Experience

Internships

2026 · 2-week engagement
AI Strategy Consultant
Planily, via Practera Industry Projects (Remote)
  • Co-authored an integrated AI strategy, marketing, and product report for Planily, a UK-based AI family organiser startup, as part of a 6-person team.
  • Proposed an AI routing framework (rules → retrieval → small model → large model → human review) balancing cost and trust as the product scales.
  • Modelled AI infrastructure costs across 50–1,000 family growth milestones, comparing paid LLM APIs against self-hosted models to identify the most cost-effective path for the pilot stage.
  • Recommended privacy-by-design as a core marketing differentiator, aligned with UK GDPR and the ICO's Children's Code, and presented a 3-year AI roadmap to startup stakeholders.
Feb – Jun 2025
AI Intern · ML Domain Lead
Cloud Counselage Pvt. Ltd. (Remote)
  • Joined the internal team as ML Domain Lead, evaluating and assessing student ML project submissions against industry standards.
  • Researched RAG techniques and LLM frameworks for potential internal knowledge tools; assessed technical feasibility for the business.
  • Designed automated workflow agents using Microsoft Power Automate and MS365, reducing manual task time by an estimated 30%.
Jun – Aug 2024
Data Analytics Intern
CSRBOX (IBM SkillsBuild, Remote)
  • Built the India Water Quality Dashboard: a Streamlit app analysing SDG 6 water data across 639 locations in 31 states with automated quality grading and Plotly visualisations.
  • EDA on surface and groundwater parameters (BOD, dissolved oxygen, pH) against WHO/CPCB thresholds to identify pollution patterns.
Jun – Aug 2023
Data Science Intern
Tanmann Foundation (Remote)
  • Explainable clinical decision-support tool using Random Forest + SHAP TreeExplainer, predicting patient disease outcomes with per-patient waterfall plot explanations.
  • Deployed as an interactive web app with global feature importance and a per-patient data explorer.
Community

Volunteering

Jul 2022 – Oct 2024
Student Volunteer
National Service Scheme (NSS)
  • Taught computer literacy to underprivileged children; organised blood donation drives and beach cleanups.
  • Participated in tree planting, awareness campaigns, and mental health surveys; supported civic engagement events.
Jul 2014 – Present
Member
Focolare Movement
  • Active member of an international movement present in 180+ countries, promoting unity and social renewal through community events, Run4Unity, and volunteering at care homes.
Education

Education

BEng Artificial Intelligence & Data Science (Hons in Cybersecurity)
University of Mumbai
Completed May 2025
First Class. ML, deep learning, data engineering, cybersecurity.
Vice-captain of my college drone club.
First Class Honours
1st Place, Our-Research HR Act
2022
1st Runner-up, HR On-Boarding
2023
1st Place, Chess
Intra & inter-college
Vice-Captain
College Drone Club
Credentials

Certifications

Google
Google Advanced Data Analytics
Dec 2024DAX0E6LSKRPT
EDAFeature EngineeringPythonML Modeling
↗ View Credential
Google
Google Business Intelligence
Dec 2024YTWL4Q1QGKUQ
BI DashboardsData PipelinesVisualisation
↗ View Credential
IBM
Databases and SQL for Data Science with Python
May 2023CR5NSY974QYW
SQLDatabasesPython
↗ View Credential
Infosys Springboard
Hands-On Feature Engineering with Python
Jul 2023
Feature EngineeringPython
Contact
Get intouch.

I'm open to opportunities in data, ML, and AI across any domain or role type. Remote, hybrid, or on-site all work for me.
If something on this site caught your eye, let’s get in touch!! ✦ Available now · UK-based · Open to relocation