Álvaro Schwiedop Souto

Álvaro
Schwiedop Souto

AI researcher & engineer

I study how models see, learn and reason. Then I build systems that put that research to work.

Get in touchRead my CV

Santiago de Compostela, Spain

Research

PHASE–TTC

From motion to time. Research paper, 2026.

Estimating time to contact by combining event-camera signals and RGB video. A causal architecture that brings motion geometry and temporal refinement together.

20.49%

lower Overall MiD than Garl-TTC35.80 vs. 45.03 on GarlTTC Test

Read the paperView code

Ranked #1 among three displayed entries on the official GarlTTC Test leaderboard as of 28 October 2026. 6,762 scored tokens with 0% scorer-defined failures. Nine-component architecture; 650 ms of causal history; a 1,955-parameter correction head. Fusion combines 20 interpretable phase, timing and support features, with grouped out-of-fold training for the fusion stages.

First page of PHASE-TTC: From Motion to Time

Causal Workspace JEPA

Mechanistic interpretability

Understanding what changes inside a language model when its answer changes. Controlled interventions and differential analysis of Qwen3-0.6B.

Methods & results

Built a 612-patch dataset with entity-disjoint splits. Activation patches transferred the donor capital answer in 50% of held-out test pairs. Averaging Jacobians across 24 training contexts reduced finite-logit-effect prediction NMSE from 0.737 to 0.354 on unseen validation entities.

Code & research

FlowTwin

Physics-informed world models

Predicting where a vessel will go next. A physics-conditioned JEPA and gradient-boosted trees, trained on public NOAA AIS trajectories.

FlowTwin model comparison on held-out vessel trips
57 future held-out trips. Mean absolute error in kilometres; lower is better.
Methods & results

A three-model ensemble reduced trajectory MAE by 11.72%, from 2.635 to 2.326 km, across 57 future held-out trips. Trip-bootstrap 95% confidence interval for improvement: 5.90–17.13%. A research evaluation on public data.

Code & results

Sonar perception

Self-supervised perception

Adapting foundation models to a different kind of vision. Small-object sonar detection with DINOv2, a learned feature pyramid and Faster R-CNN.

Sonar detection evaluation comparing frozen and fine-tuned representations in Kenai and Channel
Detection evaluation across Kenai and Channel. Individual model checkpoints.
Inspect validation examples
Annotated sonar objects missed by both compared detectors
Kenai validation: annotated objects missed by the compared detectors.
Methods & results

Controlled experiments on sparse supervision and cross-location transfer, plus a fine-detail image branch to preserve localization cues beyond the backbone patch grid.

Code & experiments

Work

From understanding the problem to delivering the system.

Trace-it

Document AI · Team project

From business policy to auditable decisions. A team-built platform that turns policies into tested Python rules, with document evidence and human escalation.

Methods & results

I owned document AI, ERP integration and deployment. The delivered evaluation reproduced 471/471 reference decisions. In a cold-cache VPS benchmark of 540 synthetic invoice PDFs, ingestion time fell 51%, from 182.1 to 89.0 seconds.

Team repository

J&A Automation

Client delivery · In production

Connecting the work on site with the numbers behind it. A multilingual platform for technician assignments, approved time and expenses, project profitability and billing.

Methods & results

Built workflows from client and project setup through approvals, immutable issued invoices, partial payments and accounting exports. Role-based financial access keeps internal economics out of customer reports. Delivered English, Spanish and Portuguese reporting, PDF/XLSX exports and production deployment.

Project repository

About

I’m an AI researcher and engineer based in Santiago de Compostela. My work connects multimodal perception, self-supervised world models and mechanistic interpretability with the practical demands of building useful software.

I work across the whole process: developing a hypothesis, designing experiments, negotiating requirements and delivering a working system. I care about both the technical details and the people who will use the result.

Full curriculum vitae

Education

Computer Engineering
Universidade de Santiago de Compostela

2024–present · Third year, 2026/27

Languages

Spanish & Galician — native
Portuguese — bilingual
English — C1, Cambridge
German — C1, TestDaF

Research

PyTorch · JEPA · Transformers · DINO
Multimodal fusion · Causal interventions
Geometric learning · Statistical evaluation

Engineering

Python · SQL · FastAPI · PostgreSQL
TypeScript · React · SvelteKit · OCR
Linux · Docker · CUDA Graphs · ONNX

Working with people

Client discovery · Requirements negotiation
Solution presentation · Technical leadership
Commercial negotiation · Deal closing

Contact

For research collaborations, engineering projects or a conversation about your next idea.

alvaro@schwiedop.com
+34 644 75 95 48
GitHubLinkedInCV / PDF