Production ML & CV architecture
Design and build AI end-to-end — from edge devices and sensors through to cloud training, serving, and monitoring. The full path, built to actually hold up in the field.
- Edge & embedded
- MLOps
- Computer vision
Senior AI/ML engineer & architect Available for assignments
I turn messy real-world signal into ML that runs in production.
Ten years building AI end-to-end — wearable sensors and edge inference, cloud ML pipelines, computer vision, and LLM systems. I ship the whole path, from the noisy input to the model serving at scale.
- Building AI
- 10+ yrs
- Disclosed funding raised
- €15M+
- Team grown from
- 1→40+
- Devices in the field
- 10,000s
01 / What I take on
Three ways teams bring me in
PoC → production
Take a model that works in a notebook and turn it into a reliable system: pipelines, evaluation, infrastructure, and the unglamorous engineering that makes it dependable at scale.
- Pipelines
- Evaluation
- Reliability
LLM & VLM systems
Structured, traceable LLM and vision-language pipelines — matching, extraction, and evaluation that stay explainable and keep per-call cost under control.
- LLM pipelines
- VLMs
- Cost & latency
Remote, part-time, invoiced through my consultancy. Most engagements run as a monthly retainer.
02 / Selected work
Signal in, systems out
Wearable AI for dairy farming, from first prototype to scale-up
One of four founders and the first employee. Designed the original neural networks that classify low-level cow behaviour from wearable sensors, wrote parts of the embedded Linux base-station stack bridging farm to cloud, ran field deployments by hand, then moved into directing a data-science team of five and the cloud ML pipelines behind it all.
accelerometer streamsbehaviour classification on 10,000s of devices
Transformer-based anomaly detection for time-series
After Datamars acquired Connecterra, I kept leading experimental projects part-time: data exploration, stakeholder alignment, and a novel transformer architecture for outlier detection in time-series sensor data.
noisy time-seriesflagged anomalies
Applied computer vision on thermal imagery
Built a web app that compares left/right body temperatures to flag potential injuries, deployed visual-understanding models (LLaVA, Qwen-VL, Grounding DINO, DensePose), and trained a custom convolutional model from scratch that outputs 2-D probability heat-maps to place regions of interest on thermal images.
thermal imagesROI heat-maps & injury flags
Privacy-first LLM matching
Built LLM pipelines that produce structured, traceable, explainable evaluations of candidates and roles — the engineering challenge being to keep evaluations deep enough to be useful while keeping per-match cost low enough to scale.
unstructured profilesexplainable matches
The same lens runs in reverse: assessing someone else’s technical design rather than building my own. Happy to support due diligence — investor or acquirer side, or a second opinion before you commit to a build.
03 / Press
Press & recognition
-
Interviewed on wearable sensors for dairy cattle.
-
Interviewed on turning raw sensor data into decisions farmers act on.
-
Connecterra won the Alpha PITCH category — 200 finalists from 2,100+ applicants. Reported by StartupJuncture.
-
Quoted on smart collars for dairy herds.
04 / Background
A few markers
- Research Co-author, peer-reviewed conference paper on quantifying dairy-cow behaviour with multi-stage SVMs (ECPLF 2017, with Aarhus University).
- Graduate study M.Sc. in Artificial Intelligence at the University of Amsterdam, supervised by Max Welling.
- Languages Spanish (native), English (C1–C2), Dutch (B1–B2), Indonesian (A1–A2).
05 / Stack
What I build with
ML & modelling
- PyTorch
- TensorFlow
- PyMC
- Computer vision
- LLMs / VLMs
- Time-series
Systems & infra
- Linux
- Kubernetes
- Docker
- Azure
- AWS
- Kafka
Data
- PostgreSQL
- MongoDB
- Event Hub
Languages & tools
- Python
- C / C++
- React
- Reflex
- uv
Fun
Outside of work
Diving
I'm a Rescue Diver with over 150 dives and I have participated in coral reef restoration programs in Thailand and Malaysia.
This tab is a work in progress; the bottleneck is me not enjoying taking pictures of myself.
Thoughts
Notes & writing
- Blog post · 2026 Demo post — placeholder title — a placeholder entry to prove out the new post layout before real writing lands here.
- Blog post · 2023 Privacy and data ownership in the age of Artificial Intelligence — anonymised data rarely stays anonymous once you cross-reference it. Why the AI systems handling customer data belong inside infrastructure you control.
More writing coming, but not promising anything either...