ZX Oncology / method review
Technical review of imaging methods: I compared definitions, thresholds and cross-cohort behaviour, and built tools to explore biomarker stability and errors.
Methodological contribution and reference reproduction.
Engineering · product · research
Jeronimo Muñoz Larreta.
Full-Stack & Machine Learning Engineer
I build full-stack software and machine learning systems. At Farmium, I work across interfaces, backend services and integrations; my machine learning work covers signals and images, from data preparation to evaluation.
Selected experience
Full-stack product development
Interfaces, services and integrations for a content platform spanning planning, media libraries, analytics and workflows.
Farmium · ProductRF signals and computer vision
Data preparation, features, classification and inference; tools to inspect and compare results.
Signals & visionMethod review and evaluation
Reproducibility, cross-cohort behaviour and threshold analysis: turning assumptions into specific checks.
ZX OncologyExperience you can
explore and inspect.
Each case connects a problem, my contribution and concrete evidence: original product software, public data or an evaluation tool.
Radio signals & drones
DroneRF / classification & evaluation
Archived results
Choose a recording and compare its reference with the archived prediction.
Explore an archived evaluationDroneRF
Waveformnormalized amplitude
Spectrogramnormalized frequency / time
ReferenceBackground
Archived outputBackground
Archived mean votes
Loading archived example…
My contribution, model and data
From signal
to evaluation.
- Problem
- Distinguish background and signals from three drone types using temporal and spectral features.
- My contribution
- Temporal and spectral feature extraction, model comparison and an HGB, MLP and Random Forest ensemble evaluated with a file-based split.
- Evidence
- Four selected public fragments, one per class, connect waveform, spectrogram, reference and the ensemble's archived prediction.
- Input per window
- 16,384 samples
- Features
- 289 temporal / spectral
- Classes
- Background + 3 drone types
- Evaluation
- File-based split
Choose a recording and compare its reference label with the ensemble output.
Data, method and attribution
DroneRF, Al-Sa’d and colleagues (2019). Signals were normalized for display; the spectrogram shows relative energy per fragment. Evaluation is offline and votes are not calibrated probabilities. Dataset and authors · CC BY 4.0.
The exploratory refit uses 211 development and 103 evaluation windows, with 51 correct predictions among those 103. The displayed selection illustrates all four classes; it does not summarize overall performance or establish field validation.
Computer vision
Thermal imaging & inspection
Inspecting correct and missed cases
Choose a crop to compare its reference, score and decision.
Inspect archived resultsInfraredSolarModules
24 × 40 · original crop enlargedInfraredSolarModules / 0
Anomaly prioritised
Archived score0.2630
Threshold0.0792
0threshold1
Dataset reference: Offline-Module
Choose an example and compare its reference with the archived decision.
My contribution, method and limits
Prioritise anomalies.
Inspect decisions.
- Problem
- Identify thermal anomalies while keeping false alarms and missed cases visible.
- My contribution
- Preprocessing and inference with ThermalNet/PyTorch, retaining the score, threshold and checkpoint.
- Evidence
- Four public crops: two correct results, a false alarm and a missed anomaly.
Model, input and evaluation
- Problem
- Prioritise anomalous thermal crops while inspecting false alarms and cases that fall below the threshold.
- My contribution
- A preprocessing and inference workflow using ThermalNet in PyTorch, checkpoint tracking and tools to compare reference labels, scores and decisions.
- Evidence
- Four original InfraredSolarModules crops with archived results: true positive, true negative, false positive and false negative.
- Input
- 24 × 40 pixels
- Model
- CNN · 3 convolutional blocks
- Output
- Anomaly / no anomaly
- Traceability
- Score, threshold and checkpoint
Compare the false alarm with the missed case. The scores fall on opposite sides of the same threshold.
Dataset and viewer scope
InfraredSolarModules — Raptor Maps (2020), MIT. Public 24×40 crops. The model shown prioritises anomaly/no-anomaly; archived scores are neither temperatures nor calibrated probabilities. Public source · MIT.
Product & systems
Farmium / platform & workflows
Frontend · Backend · APIs
Drag the nodes and traverse the dependencies of this local workflow.
Compose and traverse a workflowDAG / local computation
My contribution, architecture and APIs
Building a product,
end to end.
- Problem
- Connect accounts, media assets, planning and metrics in a single content operation.
- My contribution
- Full-stack development as co-founder: product interfaces, backend services, APIs, integrations and a visual workflow editor. I also worked on media pipelines with Python, GPU and FFmpeg.
- Evidence
- Original product interface and five navigable tools in the demo: dashboard, analytics, calendar, media library and workflows.
- Frontend
- Vue 3 · TypeScript
- Backend
- NestJS · Prisma
- Services
- Redis · S3-compatible storage
- Media
- Python · GPU · FFmpeg
The DAG on this page is a local model: drag nodes and traverse their dependency order. The full product demo is linked below.
Explore the full productView a workflow editor reference
POST /jobs
{
"type": "media.process",
"input": "asset_demo",
"callback": "webhook_demo"
}Illustrative contract for exploring an input, its status and an output event.
Research and evaluation: concrete contributions.
Technical contributions to review, reproduction and evaluation, focused on conclusions the evidence can support.
Experimental data & baselines
Import and validation, group-level aggregation and simple baselines. Units, timing and missing data are part of the contract before results are compared.
Data checks and reproducible comparisons.
Alerts, delays and missed events
I build evaluators to compare alerts by zone, type and interval, inspect idempotent duplicates, measure delivery delays and record missed events.
Independent prototype using synthetic events.
Labs
and tools.
Local-computation prototypes for exploring estimation, image comparison, search strategies and relationships between ideas.
Lab · perception and tracking
Sensor fusion
Interactive 2D model
Tracking movement
through noise.
A target moves while each sensor contributes a measurement with a different noise level. The Kalman filter combines those signals and updates its position and velocity estimate.
This model uses synthetic sensors and local computation. Explore what changes as measurements are added, removed or steered.
Toggle S1–S4, move your pointer over the field, or switch AUTO/MANUAL to steer the target.
Target, measurements and estimatelocal simulation
EstimateTargetSensors
Lab · image processing
Image comparison and sensitivity
Local computation
Change the image.
Measure the difference.
This lab explores how a perturbation changes a synthetic image and what different comparison measures capture.
It computes an SSIM measure and L∞ and L2 distances as epsilon changes or the divider moves. Processing runs in the browser.
Move the divider and adjust epsilon. The metrics describe the difference between these two images.
View the editor interface
Explore image sensitivitySSIM / L∞ / L2
Synthetic imagery and image-comparison calculations in the browser.
Lab · numerical optimisation
Search strategies
Local computation
Compare paths
to a solution.
A lab for observing how different strategies search for minima on the same function. It exposes candidates, the best value and each method’s progress.
Choose a method, change its starting point and compare the path. The guide explains each element in the field.
Lab · graphs and reasoning
Relationships & assumptions
Force simulation
Making connections
visible.
In applied research, I also contributed technical reviews, method reproduction and evaluation criteria.
This graph organizes families of critique and their connections: a way to explore structure, relationships and the questions worth making explicit.
Drag a node, inspect its label or perturb the network to see the layout reorganize.
Explore a network of ideasinteractive
Tools & areas of work
TypeScriptVue 3Node.jsNestJSAPIsWebSocketsPythonPyTorchGPUFFmpegComputer visionRF signalsExperimental evaluation
Product · software · machine learning
Let’s keep
talking.
To discuss a software or machine learning role, product development or a technical project.