Senior software engineer & computational physicist
Numerical ocean models on HPC clusters, and the systems around them. More than ten years across both.
The same coastal flood, run again and again. Each run draws an extreme sea level and a smooth local perturbation; a cell floods when its ground lies below the water. The chance of flooding at a cell is the fraction of runs that flood it, and the Monte Carlo error falls as 1/√N. Measured against an independent 25,000-run reference, it follows √(p(1−p)(1/N + 1/25,000)) — the dashed line. The country-wide flooding digital twin behind this ran a designed set of about 500 climate scenarios in each of about 50 zones, some 25,000 simulations on an HPC cluster. A designed set is not a random sample, so the 1/√N law does not describe it; this shows what repeated random runs do.
Separating slow eddies from fast internal waves. A 2-D Fourier transform turns an ocean section into a map of size and speed, where the two kinds of motion sit apart, so a filter can split them. The Ph.D. filter draws its line on the 10th internal-wave mode (waves above, eddies below); drag it to try other modes. Real eddies Doppler-shift the waves and blur the line near f, so this is the clean limit. The Ph.D. worked on about 2 TB of LLC4320 output in Eastern Boundary Currents.
Telemetry drifts before an instrument fails. A recurrent network (an LSTM) reads the recent history and outputs the chance that the instrument fails inside the next maintenance window; when it crosses a threshold, service gets booked. The traces on screen are synthetic, drawn to look like the real thing. The two figures are not: about 70 % accuracy at predicting whether a mass spectrometer fails inside a window, and an estimated 40 % less unplanned downtime.
A Lagrangian ensemble. Each of 24 oil parcels drifts at 3 % of its own wind speed plus a random walk (K = 10 m² s−¹) until it reaches the coast or 72 h pass. The fraction that land is the forecast probability, with a Monte Carlo error √(p(1−p)/N). I re-engineered the particle core of IH-TESEO to run 50× more particles; here there are 24.
Two circles added together. A charged particle in an ideal Penning trap moves as z = a eiω+t + b eiω−t: a fast cyclotron gyration around a slowly drifting guiding centre. In the lab the path never closes; in a frame turning at ω− the drift vanishes and the guiding centre sits still. The M.Sc. thesis took this to the quantum case, with an added rotating field, and its coherent states.
Spectral clustering of a tiny image. The 1,440 pixels are linked by how alike their colour and position are. The eigenvectors of the normalised graph Laplacian L = I − D−½WD−½ with the smallest eigenvalues encode the regions, and k-means on their rows draws the segments. The B.Eng. thesis replaced the eigenvector step with a quantum algorithm; the classical version is what runs here.
Frequencies are masses. Ions in a magnetic field orbit at ω = qB/m, so heavier ions orbit slower. Each induces a tiny image current on the detection plates; the sum is a decaying transient whose Fourier transform is a mass spectrum. A longer record resolves finer peaks: resolution grows with record length. Watch the overlapping lumps separate.
25,000 independent scenarios, run in a background thread while you watch. A cell’s chance of flooding is the fraction of runs that flood it. The plotted error is measured: 10,000 further independent runs, cut into blocks of N and compared with the 25,000-run map.
Terrain, river channel and shoreline are generated from value noise. The water level is Gumbel-distributed (extreme-value) and a smooth random perturbation stands in for local surge and wave variability. Heights are in arbitrary units.
A cell floods if its ground is below the water: a bathtub rule with no hydrodynamics. The real project ran full numerical simulations on a cluster, on a designed set of scenarios rather than random draws; this shows the statistics of repeated random runs, not that study’s error.
A 2-D FFT of the field, a mask below the boundary curve, and the inverse FFT. The bright spots in the spectrum are the FFT’s own output; the curves through them are the theoretical dispersion relation.
Three slowly drifting eddies, plus waves on vertical modes 1 to 3 in both directions. Frequencies follow ω² = f² + (c/n)²k² for modes n = 1, 2, 3. The boundary starts on the mode-10 curve.
Linear, hydrostatic waves in a quiet background. Real eddies Doppler-shift them and blur the line near f, which is the hard part of the real problem.
Only the animation. The curves come from a seeded random stream; no model is running in your browser.
≈70 % accuracy and ≈40 % less unplanned downtime (estimated) are the project’s own numbers. Nothing else on this tile is.
Temperature, vacuum pressure and ion current with a drift before the failure, and a smooth probability curve. The real telemetry and model are not shown, and the days of warning are illustrative.
24 parcels, each with its own wind and a random walk, stepped every 30 minutes for 72 hours. The landing probability and its error are counted from the parcels.
A drawn coastline, one release point, and a wind of 12 m/s with ±15 % speed and ±8° heading spread. The coastline is not a real chart.
Surface drift and diffusion only. The production model also weathers the oil and uses ocean currents; neither is here.
The analytic motion evaluated at 5,200 points, and the transform z′ = z e−iΩt for the right-hand panel as Ω sweeps from 0 to ω−.
a = 1, b = 0.58, ω+ = 1, ω− = 0.07 in arbitrary units. In a real trap ω+/ω− is 10³ or more, which would make the drift invisible at this scale.
Classical motion in an ideal trap, without the rotating transverse field the thesis adds. The glowing packet is a stand-in drawn at 10 % of the orbit; a real coherent state is about 10 nm across.
The 1,440 × 1,440 Gaussian affinity, the normalised matrix, subspace iteration with Rayleigh–Ritz for its top eigenvectors, row normalisation and k-means with k = 4, in a background thread.
A 48 × 30 image drawn from a sunset sky, a sun, glittering sea and a dark headland, plus 2 % noise.
A very small image with a dense matrix. Real images need sparse graphs and iterative eigensolvers, and the quantum algorithm is not simulated here.
The 4,096-sample transient is generated once; every frame, the Fourier transform of its first n samples is recomputed, so the peak widths are the real consequence of the record length.
Five species (m/z 100, 125, 150, 200, 250) with illustrative amplitudes, a common decay time and a little noise. The ions are not a real sample.
The ideal cyclotron frequency, with the trapping field and space charge neglected. A real Penning trap measures the shifted reduced cyclotron frequency ω+ < ωc.
01 / The bridge
Scientific computing and distributed backend systems share fundamental requirements: numerical correctness, throughput under load, and reproducible execution. I have spent more than ten years working across both domains, applying the same rigor whether optimising Fortran and OpenMP kernels on an HPC cluster or operating Go microservices under high traffic.
01
When re-engineering IH-TESEO's particle core, I implemented automated regression tests for weathering processes and integrated them into GitHub Actions, ensuring numerical integrity is verified on every commit.
02
Whether handling thousands of ticket purchases per minute, hundreds of asset validations per second, tens of gigabytes of ocean data daily, or 25k climate-scenario simulations on a cluster, reliable throughput comes from identifying bottlenecks and profiling early.
03
Containerising numerical models like ROMS, Delft3D, and SFINCS ensures simulations are portable and results remain defensible over time. Consistent environments make deployments predictable and scientific results reliable.
04
Experience taking systems from architecture and development through Docker deployment, CI/CD, and operational maintenance—including unattended data pipelines, alerting, and internal support tools.
02 / Results, measured
Selected figures from systems I have designed, built, and operated across high-performance computing, operational data engineering, and backend infrastructure.
50×
more particles simulated
IH-TESEO Lagrangian transport, re-engineered Fortran/OpenMP core — with weathering physics fully validated and no loss of accuracy.
4× faster
execution speedup
Same model, up to 4× faster, plus netCDF I/O so output interoperates with standard ocean-data tooling.
100GB/day
ingested unattended
Operational forecast system producing daily ocean-current and sea-level predictions for the Cantabrian coast and Santander Bay.
25k
HPC scenarios executed
A country-wide coastal flooding digital twin — ~500 climate-change scenarios across ~50 estuary-like zones, run entirely on an HPC SLURM cluster, with GeoPandas driving scenario generation.
~2TB
simulation output processed
Doctoral ETL pipeline over the LLC4320 high-resolution global ocean run, feeding a spectral wave–balanced filter.
O(1k)/min
peak users handled
Live-event ticketing infrastructure supported thousands of purchases per minute while validating hundreds of asset ownerships per second.
~50k
accounts onboarded
Verifiable credentials issued and validated across thousands of devices via Go microservices (REST + gRPC).
<1s
eligibility check latency
End to end, and everything is inside it: cryptographic verification, fetching the asset state from the blockchain, and the configurable rules engine's own validation — fast enough to keep gate statistics near-real-time.
5k
users via third-party SDK
Glyph's open-source SDK brought in 5,000 users beyond the 2,000 internal adopters it reached in its first month.
40%
less unplanned downtime
Estimated improvement from an LSTM model predicting mass-spectrometer failure inside a scheduled maintenance window.
03 / Outstanding work
Case studies outlining problem constraints, architecture decisions, and operational outcomes.
Case 01 · IHCantabria
Fortran/OpenMP re-engineering, netCDF I/O, validated weathering physics and a CI suite guarding the numerics.
Case 05 · tokenproof
Go microservices issuing and verifying signed credentials, a configurable eligibility rules engine, and on-device ECDSA for Android and iOS.
Case 04 · Yuga Labs
Owned end to end and open-sourced: a widget plus an SDK that let third-party developers integrate multi-chain EVM accounts into their applications.
04 / The science
Interactive browser implementations of research work from my doctoral, master's, and engineering theses — demonstrating wave–vortex decomposition, Penning trap dynamics, and spectral clustering running in real time.
Ph.D. · CICESE
2 TB of a global ocean simulation, filtered in wavelength–frequency space to split gravity waves from geostrophic flow. Drag the latitude slider and watch the partition move with the Coriolis frequency.
M.Sc. · CINVESTAV-IPN
Coherent states, ladder operators and eigenvalues, derived analytically and checked numerically. The demo integrates the Lorentz equations live so you can destabilise the trap yourself.
B.Eng. · UPIITA-IPN
A quantum variant of spectral clustering with theoretically constant runtime for the whole algorithm, under idealised assumptions. The demo builds the graph Laplacian, extracts the Fiedler vector, and shows k-means losing to it on the same points.
05 / Sources and provenance
Each demonstration says what is computed live, what is synthetic and what is simplified. Entries marked † are standard textbooks or classic papers, cited without a link. Colours follow the Okabe–Ito colour-blind-safe set, with perceptual maps (mako, inferno) for continuous fields.
06 / Availability
This is most useful where a hard problem needs owning end to end — architecture, implementation, deployment and the day-2 reality — and especially where scientific or numerical correctness is part of the product. Based in Santander, working remotely, with hours that overlap both Europe and the Americas.