Daniel Huencho Mardini

Research Assistant — Grand Challenges Living Lab

MSc AI for Sustainable Development | UCL

2026-01-26

Technical Knowledge

2011–2018 Civil Engineering
PUC Chile

2019–2020 MSc Statistics
PUC Chile

2025 MSc AI for Sustainable
Development — UCL

Probabilistic Modelling

• Deep Kernel Learning (DKL) for dissertation
• Bayesian modelling & uncertainty quantification
• Gaussian Process Summer School (gpss.cc)

Tools & Frameworks

• Python, PyTorch, GPyTorch, scikit-learn
• Time series analysis & forecasting
• Probabilistic ML pipelines

• Walk through the education path: started in Industrial/Civil Engineering at PUC Chile — strong mathematical foundation in calculus, linear algebra, probability, operational research.
• Then MSc in Statistics at PUC Chile — ranked 2nd in cohort, GPA 3.75/4.0. Thesis on multinomial logit models. Modules in Bayesian Statistics, Time Series, GLMs, ML Applications.
• Currently at UCL doing MSc AI for Sustainable Development — modules in Probabilistic Modelling, Deep Representations and Learning.
• For the dissertation: exploring Deep Kernel Learning (DKL) — combines deep neural networks with Gaussian Process priors. Interested in the trade-offs of uncertainty quantification.
• Actively studying from GP Summer School (gpss.cc) to deepen understanding of kernel methods.

Passion for Applicable Knowledge

I defined my own dissertation project — I connected Professor Carmine Galasso (IRDR) with Professor Benjamin Guedj (Computer Science) to create a cross-departmental research project combining structural engineering with probabilistic ML.

Research connections — Attended Henry Moss's talk "Experimental Design in the Age of Generative Models" (constraints in optimisation of generative models). Met PhD student Daniel Giles there. Worked with Max Harris on next-token generation for customisation (Knowledge Management lecture project).

"We are enhancing models for capability, not for intelligence" — Oxford Professor

This is why I am drawn to compositional kernel methods and Gaussian Processes — interpretable, uncertainty-aware, and decomposable into meaningful components.

• Key message: I don’t just study what I’m told — I actively seek and create research opportunities.
• Dissertation story: I identified the gap between structural risk (Galasso’s domain) and probabilistic ML (Guedj’s domain). I approached both professors independently and proposed the project that became my dissertation.
• Henry Moss’s talk was about how we need constraints and structure in the optimisation of generative models — not just scaling. This resonated with my interest in interpretability.
• Daniel Giles is a PhD student I connected with at these events — building my research network.
• Max Harris collaboration on next-token generation was a lecture project for Knowledge Management and Education module.
• The Oxford professor quote captures why I care about GPs: we need models where we understand what the components mean, not just models that predict well.
• Compositional kernels and GPs are not just a tool for me — they are a core research interest directly aligned with this role.

Engineering & Collaboration

Production ML at Scale

• SCADA energy data pipeline at Metro de Santiago
• AWS: S3, Glue, Athena, Lambda, EC2, CodePipeline
• Anomaly detection on real sensor data
• Predictive maintenance for ticket machines
• CI/CD, agile framework, testing discipline

Interdisciplinary Teamwork

• Led team of 5 data scientists & engineers
• Bridged engineering, operations, and executive teams
• Value diverse perspectives for real-world solutions

Personal Stake in UCL Energy

• SRA at Ramsay Hall — I see UCL's energy usage patterns daily
• This project is personal, not just professional

• Left column: Talk about the Metro de Santiago experience as Head of Data Science. Real engineering, not just notebooks.
• The SCADA energy pipeline handled high-frequency sensor data across overlapping electrical zones. Had to deal with data quality, spatial-temporal challenges, and production reliability.
• AWS stack: S3 for storage, Glue for ETL, Athena for queries, Lambda for serverless, EC2 for compute, CodePipeline for CI/CD. Real production architecture, not prototypes.
• Working with sensors requires caution and robustness — data drift detection, anomaly detection, thinking about the final user.
• Right column: Emphasise the interdisciplinary nature. At Metro, I worked with operations teams, electrical engineers, executives — not just data people.
• The SRA (Student Residence Adviser) point is key: I live in UCL accommodation, I see the energy usage first-hand. This makes the Grand Challenges Living Lab project personal.
• Mention collaboration style: connected Galasso with Guedj, worked with Nicole Paul, Daniel Giles, Chloe — I build bridges between disciplines.

I Built Exactly What You Need

UNICON GP Demo — Compositional Gaussian Process modelling on university energy data:

• 6-component additive kernel (trend, annual, weekly, weather, calendar, events)
• 8 building categories, anomaly detection, cross-building comparison
• Kernel decomposition into interpretable components

View Live Demo →

8 Building Categories

6 Kernel Components

16 Visualisations

GP Compositional

daniel.mardini.25@ucl.ac.uk  •  danielhuencho.com  •  github.com/choka30

• This slide is the proof. Don’t just talk about skills — show the work.
• The UNICON demo is a full compositional GP modelling pipeline on real university energy data (La Trobe University, Australia). This is directly analogous to what the Grand Challenges Living Lab project needs for UCL estates.
• Walk through the 6-component kernel: trend (long-term baseline), annual seasonality (climate-driven), weekly periodicity (occupancy), weather response (nonlinear), calendar effects (academic year), events (interventions like COVID, LED installations).
• 8 building categories modelled independently — enables cross-building comparison of how GPs capture different building physics. This is exactly the “generalisation across locations” the role description asks for.
• Anomaly detection using standardised GP residuals — higher use than expected from similar locations.
• The demo is live at danielhuencho.com — encourage them to explore it.
• Close with contact information and express genuine enthusiasm for the role.