Carotid ultrasound segmentation
Deep learning to automate the segmentation of the carotid artery and the classification of atheroma plaque in ultrasound scans, built side by side with a PhD researcher.
- Keras
- TensorFlow
- Semantic segmentation
Freelance Data Scientist & AI Engineer
I build machine learning that does real work.
Physics first. Then eight years of shipping models. Every project below opens up to show how it was built.
fig. 0 — Bosons in a 2D harmonic trap, the system from my first paper (Phys. Rev. A 96, 043614, 2017).
01 · Computer vision
Teach machines to see what your experts see.
Detection, segmentation and measurement in medical, biological and industrial imagery.
fig. 1 — A carotid bifurcation. Highlighted: the bulb, where plaque tends to build up.
Cell detection & morphology
A lab wanted to move from looking at microscopy images to measuring them. Every cell is detected, segmented and described by its shape, and the structures cells form are mapped across the image grid.
→ Qualitative observations became reproducible numbers.
- YOLO
- Cellpose
- StarDist
- CellProfiler
- OpenCV
fig. 2 — Segmented cells, nuclei and detection boxes.
02 · Documents & LLMs
Read thousands of pages so your team doesn't have to.
Pipelines that extract, check and structure information from documents, including when the data can't leave the building.
Legal compliance checker for PDFs
Auditors reviewed supplier documents by hand: slow, repetitive, and inconsistent from one auditor to the next. Every supplier rebranded the official template differently. The system checks each PDF against 50–100 legal rules, from layout and icons to whether each section says what the law requires.
→ From about 15 minutes of manual review per document to under a minute, with the same criteria for everyone.
- Tesseract OCR
- CLIP
- Phi-4
- OpenCV
- On-premise
- CPU only
- No external APIs
- Deterministic
fig. 3 — A supplier document, as the auditors received it.
03 · Risk & fraud scoring
Score risk before it turns into a loss.
End-to-end scoring systems the business relies on: from raw records to a score, a decision and a report.
Booking risk model
Truvi protects hosts and guests in short-term rentals. I own the model that scores every booking so the business can decide how to protect each stay. I took over an existing model and improved it, and I'm making it scale with the business.
→ Better predictions that are easier to interpret for the business, running in production.
- Python
- SQL
- Machine learning
- MLOps
- Reporting
fig. 4 — A stream of bookings. Highlighted: the few that need extra protection.
Bank transaction classifier
UNNAX aggregated bank accounts so lenders could assess their customers. I designed a production classifier, and kept improving it for two years, that sorts short, noisy, multilingual transaction descriptions into about 10 categories and 30 subcategories. It later went from Spain to Mexico.
→ Customer profiles (risk factors, ability to pay) built on top of the categories.
- Python
- R
- scikit-learn
- SQL
fig. 5 — One stream of transactions, sorted into categories.
04 · Forecasting & anomaly detection
Spot the failure before it happens.
Time series models on sensor data: early warnings, real-time quality checks, and less time lost to breakdowns.
Predictive maintenance for solar trackers
Motors and batteries in solar trackers fail, and a spare part can take weeks to reach a plant. Using multivariate sensor data from more than a thousand trackers worldwide, the system detects abnormal behaviour and classifies each component's health.
→ About 30 days of warning, enough time to get the spare part on site.
- Python
- scikit-learn
- Autoencoders
- Time series
fig. 6 — A field of solar trackers. Highlighted: the one about to fail.
Real-time quality control in laser manufacturing
In a subtractive laser process, every second spent on a part that's already defective is wasted. The model reads the laser's power signal as it streams and flags bad parts mid-process, before the part is finished.
→ Millisecond inference over millions of data points, running live on the client's experimental production line.
- Python
- Autoencoders
- Streaming
- scikit-learn
fig. 7 — A part cut layer by layer. Highlighted: the laser.
05 · Measurement at scale
Numbers an entire industry can trust.
Audited data products where every change of methodology has to be explained.
Audience currencies for Sweden and Spain
An audience currency is the number broadcasters, publishers and advertisers trade on. We combined a panel of thousands of households with large daily volumes of TV and online events into one currency: Total Video in Sweden and Digital Audience Measurement (DAM) in Spain.
→ Reliable, audited deliveries.
- Python
- R
- AWS
- SQL
- Terraform
- GitLab CI
fig. 8 — TV, laptop, tablet and phone, merged into one measure.
06 · Scientific ML
Machine learning for hard science.
From quantum physics to drug discovery: research-grade work, honestly benchmarked.
Virtual screening for drug discovery
Finding, by computation, the molecules likely to bind a target protein. I implemented a graph convolutional autoencoder that compares candidates with known ligands, plus an ML docking-quality score, as the baseline for a team developing quantum-inspired algorithms.
→ A fair benchmark: ML and quantum-inspired methods came out neck and neck, each winning some cases.
- PyTorch
- Graph neural networks
- Autoencoders
- Docking
fig. 9 — A candidate molecule.
07 · AI adoption for teams
Get your team working with AI, not just chatting with it.
Hands-on productivity work, so analysts and engineers use AI across their real workflow.
From chat window to Claude Code
Truvi's data team only used AI through a chat window. I taught them to work with Claude Code on their own workstations and to use it across several phases of development, not just for one-off questions.
- Claude Code
- Productivity
- Workshops
fig. 10 — A team, connected.
The path so far
From quantum physics to production ML.
- 2026 –Truvi (UK, remote)Freelance data scientist
- 2024 – 2026Izertis (ES, remote)Senior Data Scientist, Emerging Technologies
- 2021 – 2024GfK (DE, remote)Data Scientist
- 2019 – 2021UNNAX (Barcelona)Data Scientist
- 2018 – 2019AKKA Technologies (Toulouse)Data science consultant
- 2017 – 2018University of BarcelonaMSc Fundamentals of Data Science
- 2016300.000 km/sPython developer intern
- 2013 – 2017University of BarcelonaBSc Physics · co-author, Phys. Rev. A
fig. 11 — 2013 to today, and a dotted line for what's next.
Let's talk
Something that should be faster, smarter or automatic?
Tell me about the problem. I'll tell you honestly what fits: machine learning, an AI workflow, or plain automation with no AI at all, and what it would take to get it running.
hello@enricsarle.com · I usually reply within a day.
fig. ∞ — Same particles, one last shape.