Jordi Laguarta Soler
PhD candidate, MIT EECSEarth Intelligence Lab
I build multimodal ML systems end to end for domains where expert labels are scarce: foundation models as annotators, self-supervised pretraining, explainable predictions, and national‑scale inference evaluated under distribution shift. At MIT and Microsoft Research, I have applied them to high‑stakes problems in health and food security: screening for COVID‑19 and Alzheimer's from coughs and speech, and mapping crops across entire countries.
I lead research on evaluation science: whether geospatial benchmarks can reliably tell foundation models apart. I have mentored student researchers for over three years.
PhD expected February 2027 · seeking research and applied scientist roles; open to forward‑deployed and research‑product work.
- VLM-as-annotator
- Cross-view pseudo‑labelling
- Self-supervised multimodal pretraining
- Cross-modal distillation
- Country-scale inference
- High-throughput data streaming
- Distribution shift
- Benchmark reliability
Work
-
A 10 m crop map of India, supervised by a vision-language model
I prompt a VLM to label street‑level images, then distil it into a lightweight classifier on foundation‑model embeddings that maps crops at national scale.
-
Acoustic biomarkers for COVID‑19 and Alzheimer's screening
Acoustic-biomarker encoders pretrained on separate audio tasks transfer to COVID‑19 screening from coughs and Alzheimer's screening from speech, with per-patient saliency maps. I was first author on both papers.
-
A multimodal foundation model for satellite time series
I designed the encoder architecture, built the data‑streaming pipeline and ran distributed multi‑GPU pretraining on radar, optical and land‑cover time series.
-
Which Earth-observation benchmarks can tell foundation models apart?
I lead an analysis of published geospatial foundation-model results that tests, benchmark by benchmark, whether the gaps between models exceed seed noise.
-
A national crop map of Thailand from Street View and zero-shot VLM labels
I replaced expert annotation with zero-shot GPT‑4V labels as the training signal, and the map comes close to the expert‑supervised one.
Experience
Hover the timeline for details- 2022–nowMIT EECS · PhD candidate, Earth Intelligence Lab (advisor: Sherrie Wang)
First author on the national crop-map papers for Thailand (AAAI 2024) and India (in preparation), which use vision-language models to label street‑level imagery. Leading research on benchmark reliability for geospatial foundation models. Mentored three student researchers over three years.
- 2024Microsoft Research · Research intern, Seattle
Designed a self-supervised multimodal foundation model for radar, optical and land‑cover time series; built its 20 TB data‑streaming pipeline and ran distributed multi‑GPU pretraining.
- 2021–22EY-Parthenon · Strategy consulting intern, Geneva
Built a playbook of technology-led sustainability strategies for consumer-goods, retail and agribusiness companies.
- 2020–21MIT Auto-ID Lab · Research assistant
First author on the COVID‑19 and Alzheimer's screening papers. Set up clinical data collaborations with the Government of Catalonia's Department of Health (Spain) and the Universidad de Guadalajara (Mexico) for forced‑cough recordings from PCR‑tested patients.
- 2018–19Nutrino Health · ADLA · Applied ML at start-ups
Led the data-science team building a computer-vision recommendation chatbot at ADLA; improved a food‑image classifier with weakly labelled data at Nutrino Health (later acquired by Medtronic).
- EducationMIT PhD EECS · Cornell Tech MEng CS · Imperial College London BEng EIE (machine learning)
Publications
All publications →- An open 10-meter multi-crop map of India from street-level imagery and vision-language models in preparation
- Combining deep learning and street view imagery to map smallholder crop types AAAI 2024
- COVID‑19 artificial intelligence diagnosis using only cough recordings IEEE OJEMB 2020 · 700+ citations
- Longitudinal speech biomarkers for automated Alzheimer's detection Frontiers in CS 2021